<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Mikhail Drozdov on Medium]]></title>
        <description><![CDATA[Stories by Mikhail Drozdov on Medium]]></description>
        <link>https://medium.com/@casinokrisa?source=rss-a72f34cd294c------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*NLHsEvABZ9RrmIXMrpPGvQ.png</url>
            <title>Stories by Mikhail Drozdov on Medium</title>
            <link>https://medium.com/@casinokrisa?source=rss-a72f34cd294c------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 22:06:22 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@casinokrisa/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Outdated retrieval can overturn correct answers]]></title>
            <link>https://medium.com/@casinokrisa/outdated-retrieval-can-overturn-correct-answers-8a7f09cee427?source=rss-a72f34cd294c------2</link>
            <guid isPermaLink="false">https://medium.com/p/8a7f09cee427</guid>
            <category><![CDATA[search]]></category>
            <category><![CDATA[seo]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Mikhail Drozdov]]></dc:creator>
            <pubDate>Mon, 28 Sep 2026 09:07:55 GMT</pubDate>
            <atom:updated>2026-09-28T09:07:55.571Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Canonical: </em><a href="https://casinokrisa.com/blog/stale-document-poisoning-retrieval-validity-2026-09"><em>https://casinokrisa.com/blog/stale-document-poisoning-retrieval-validity-2026-09</em></a></p><p>On 28 September 2026, Ahmed, Galke, Poech, and Röttger posted <a href="https://arxiv.org/abs/2609.31342">Stale-Document Poisoning</a> (arXiv:2609.31342). Their finding cuts against the usual RAG sales pitch: retrieval is supposed to fix stale weights, yet <strong>a single authentic but outdated document can flip answers the same model already gets right without retrieval</strong>.</p><p>This is not adversarial poisoning. The documents were once correct. They become harmful when the conditions that made them valid change.</p><h3>What they measured</h3><p>Poisoning is counted only when the model is correct without retrieval and wrong after one outdated document is inserted. Across 317 source-verified reversals (medicine, law, software, platform policy):</p><ul><li>Neutral outdated retrieval flips roughly <strong>30%</strong> of Llama and <strong>37%</strong> of Qwen answers that were previously correct.</li><li>Explicit “follow the dated documentation” instructions raise those rates to about <strong>66%</strong> and <strong>75%</strong>.</li><li>Matched <strong>up-to-date</strong> evidence is followed in about <strong>97–100%</strong> of trials.</li><li>Cross-domain instructed poisoning spans about <strong>17–91%</strong> (software lowest, medicine highest).</li></ul><p>So models can use retrieval. They do not reliably refuse superseded retrieval. That is a failure of <strong>selective epistemic trust</strong>.</p><h3>Dates are not validity</h3><p>On 50 transitions, holding the evidence fixed and changing only the evaluation date produces only modest adaptation. Add an explicit validity boundary (when the old claim stops applying), and 70B/72B-class models switch almost perfectly. Causal patches at the evaluation-date position move the answer; matched patches at the unchanged source date do not.</p><p>A recency-aware re-ranker cuts poisoning by about <strong>4.6–10.0 percentage points</strong> only when dates are accurate. Wrong dates can make it worse.</p><h3>Why this matters for indexing systems</h3><p>Crawl–index–serve answers whether a URL is known and eligible. Validity answers whether the claim still governs. Publication date, crawl timestamp, and lastmod are weak proxies for supersession.</p><p>Google’s public bar for AI features on Search remains index + snippet eligibility. Necessary, not sufficient, for <strong>timely</strong> supporting links. A superseded page that stays indexed can still enter any answer stack that ranks relevance without applicability.</p><p>Product defaults that say “always follow the retrieved doc” amplify the failure the paper measures.</p><h3>Limits</h3><p>Secondary reading of one arXiv preprint. No local replication. Cite the paper’s tables before treating any percentage as universal.</p><h3>Takeaway</h3><p>Reliable RAG needs selective trust: not only what evidence says, but whether it still applies. Treat effective dates and supersession as first-class index metadata.</p><p>— Mikhail Drozdov, AI Search &amp; Indexing Systems Researcher · <a href="https://casinokrisa.com/person/mikhail-drozdov">https://casinokrisa.com/person/mikhail-drozdov</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8a7f09cee427" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Crawl Stats gaps are not Page Indexing no-backfill]]></title>
            <link>https://medium.com/@casinokrisa/crawl-stats-gaps-are-not-page-indexing-no-backfill-1c23e19ce040?source=rss-a72f34cd294c------2</link>
            <guid isPermaLink="false">https://medium.com/p/1c23e19ce040</guid>
            <category><![CDATA[seo]]></category>
            <category><![CDATA[search]]></category>
            <category><![CDATA[measurement]]></category>
            <dc:creator><![CDATA[Mikhail Drozdov]]></dc:creator>
            <pubDate>Mon, 21 Sep 2026 09:06:57 GMT</pubDate>
            <atom:updated>2026-09-21T09:06:57.257Z</atom:updated>
            <content:encoded><![CDATA[<p>Canonical: <a href="https://casinokrisa.com/blog/gsc-crawl-stats-gap-vs-page-indexing-2026-09">https://casinokrisa.com/blog/gsc-crawl-stats-gap-vs-page-indexing-2026-09</a></p><p>On 20 September 2026 Search Engine Roundtable reported a universal Google Search Console Crawl Stats hole for 15 September (some time zones may show 16 September). Industry history says those Crawl Stats gaps usually get restored. That is not the same recovery class as John Mueller’s 11 September statement that Google does not back-fill Page Indexing data for the June delay hole.</p><p>Two sensors, two expectations</p><p>Crawl Stats — Recent hole: 15 Sep 2026 day missing (~20 Sep notice). Recovery class: often restorable; wait; keep exports; do not treat blank as crawl stop.</p><p>Page Indexing — Recent hole: June 2026 multi-day blanks (~11 Sep notice). Recovery class: not promised to restore (Mueller: no back-fill of indexing data).</p><p>Crawl Stats measures fetch history (requests, responses, host availability). Page Indexing measures coverage status over time. A blank day on one chart does not inherit the other chart’s repair policy.</p><p>Wrong moves</p><p>- Rewriting crawl-budget narratives from one missing Crawl Stats day without Host status or server-log corroboration.<br>- Interpolating indexed URL counts across Page Indexing holes.<br>- Treating either shared, multi-property hole as proof your site alone was de-crawled or deindexed.</p><p>Check instead</p><p>For Crawl Stats blanks: Host status, response table, robots.txt availability, your access logs.<br>For Page Indexing blanks or blips: URL Inspection and live Search on stable URLs; do not invent the missing series.</p><p>Full tables, timeline crosswalk, and limits: <a href="https://casinokrisa.com/blog/gsc-crawl-stats-gap-vs-page-indexing-2026-09">https://casinokrisa.com/blog/gsc-crawl-stats-gap-vs-page-indexing-2026-09</a></p><p>— <br>Mikhail Drozdov · AI Search &amp; Indexing Systems Researcher · <a href="https://casinokrisa.com/person/mikhail-drozdov">https://casinokrisa.com/person/mikhail-drozdov</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1c23e19ce040" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Search Console does not backfill Page Indexing gaps]]></title>
            <link>https://medium.com/@casinokrisa/search-console-does-not-backfill-page-indexing-gaps-9c17461b44f4?source=rss-a72f34cd294c------2</link>
            <guid isPermaLink="false">https://medium.com/p/9c17461b44f4</guid>
            <category><![CDATA[seo]]></category>
            <category><![CDATA[indexing]]></category>
            <category><![CDATA[search]]></category>
            <category><![CDATA[google-search-console]]></category>
            <category><![CDATA[measurement]]></category>
            <dc:creator><![CDATA[Mikhail Drozdov]]></dc:creator>
            <pubDate>Mon, 14 Sep 2026 09:22:18 GMT</pubDate>
            <atom:updated>2026-09-14T09:22:18.189Z</atom:updated>
            <content:encoded><![CDATA[<p>On 11 Sep 2026 John Mueller said Google does not back-fill Page Indexing data for the June delay hole. Separately, 8–9 Sep saw a short false &quot;Crawled - currently not indexed&quot; wave. Blank days ≠ mass deindexing; blips ≠ mass removal.</p><p>Canonical: https://casinokrisa.com/blog/gsc-page-indexing-no-backfill-2026-09</p><p>Two different Page Indexing failure modes landed in the same week of industry reporting.</p><p>Gap (noticed 11 Sep 2026). Multi-day blanks for June 2026 appeared on the Page Indexing chart across properties. Search Engine Land and Search Engine Roundtable reported John Mueller&#39;s reply: the hole is likely from the June period when the report was not updated; there is no data for that period; we don&#39;t back-fill indexing data.</p><p>Blip (8–9 Sep 2026). For a short overnight window, many still-indexed URLs — including homepages — briefly showed as &quot;Crawled - currently not indexed.&quot; Mueller called it a short blip that was resolved quickly, and noted that reporting can differ while indexing still works.</p><h2>Operational distinction</h2><p>| Mode | What it is | What not to do |</p><p>| --- | --- | --- |</p><p>| Blank day | Missing telemetry Google does not promise to restore | Interpolate across the hole; read the gap as a quieter site |</p><p>| Status blip | Temporary wrong label on still-indexed URLs | Mass-resubmit or sitemap panic without URL Inspection / live check |</p><p>Page Indexing is lossy telemetry, not a continuous audit tape of the index. The mid-June delay (documented by SEL on 26 Jun, updated 3 Jul through ~29 Jun) already showed that fresh chart points can stall; September&#39;s missing June days show that some delayed windows may never get restorable points.</p><p>Full method, facts table, timeline, and limits: https://casinokrisa.com/blog/gsc-page-indexing-no-backfill-2026-09</p><p>— Mikhail Drozdov, AI Search &amp; Indexing Systems Researcher (Casinokrisa). https://casinokrisa.com/person/mikhail-drozdov</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9c17461b44f4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Dynamic AI Overview expansion does not rewrite the impression rule]]></title>
            <link>https://medium.com/@casinokrisa/dynamic-ai-overview-expansion-does-not-rewrite-the-impression-rule-c758ffa8801b?source=rss-a72f34cd294c------2</link>
            <guid isPermaLink="false">https://medium.com/p/c758ffa8801b</guid>
            <category><![CDATA[ai-search]]></category>
            <category><![CDATA[seo]]></category>
            <category><![CDATA[search-console]]></category>
            <category><![CDATA[indexing]]></category>
            <category><![CDATA[google]]></category>
            <dc:creator><![CDATA[Mikhail Drozdov]]></dc:creator>
            <pubDate>Mon, 07 Sep 2026 09:51:46 GMT</pubDate>
            <atom:updated>2026-09-07T09:51:46.286Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*q2RP1sR04SPf-R6fJ_7LnA.png" /></figure><p>Google confirmed on 28 August 2026 that AI Overviews may dynamically expand for some queries. Search Console still counts AI Overview impressions under the existing scroll-or-expand rule. Expansion is SERP chrome, not a new impression metric.</p><p>Canonical: https://casinokrisa.com/blog/aio-dynamic-expand-impression-rules-2026–08</p><p>On 28 August 2026 Google confirmed that, for some queries, AI Overviews may open to a full response without a “Show more” click, push classic results further down, and cancel that expansion if the user has already started scrolling. That is a presentation change on the results page.</p><p>It is not, on the published Help Center pages, a rewrite of Search Console’s impression methodology.</p><p>What Console already said</p><p>Help answer 7042828 still governs AI Overview counting. An impression means a link was seen (or potentially seen). For expanding widgets, the usual bar is scroll-into-view or expand. For AI Overviews specifically, the English Help Center wording says standard impression rules apply and that the link must be scrolled or expanded into view to count. The Overview occupies one position; every supporting link inherits that position.</p><p>The Generative AI performance report remains an impression ledger: pages, countries, devices, dates. Still no Queries, Clicks, CTR, or Position columns on that dedicated surface.</p><p>What the confirmation adds</p><p>Press reports (Search Engine Land, The Verge) quote Google that expansion happens where systems judge a longer answer most useful, and that early scrolling cancels expansion so people do not lose their place. Neither report documents a new Console dimension or a revised Help Center definition.</p><p>So the mechanical reading is narrow:</p><ol><li>Links that used to sit behind “Show more” can enter the expanded-into-view state without that click when expansion actually fires and the link becomes visible under the existing rule.</li><li>2. Links already in the compact viewport were already on the current results surface.</li><li>3. Classic blue links below the Overview are a different element. Shared Overview position is not a chart of how far down organic results moved.</li></ol><p>What you still cannot conclude</p><p>You still cannot reverse an impression total into the query that triggered the Overview, prove your URL authored a sentence, or read an expand-state split from the Generative AI report. Treat third-party CTR studies as industry measurement, not Google documentation.</p><p>Eligibility for supporting links still starts at indexed + snippet-eligible Search pages. Dynamic expansion does not invent a separate AI index.</p><p>Full method, table, sources, and limits: https://casinokrisa.com/blog/aio-dynamic-expand-impression-rules-2026–08</p><p>— Mikhail Drozdov, AI Search &amp; Indexing Systems Researcher (Casinokrisa). https://casinokrisa.com/person/mikhail-drozdov</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c758ffa8801b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Generative AI performance report is an impression ledger, not a query index]]></title>
            <link>https://medium.com/@casinokrisa/the-generative-ai-performance-report-is-an-impression-ledger-not-a-query-index-a38b70894969?source=rss-a72f34cd294c------2</link>
            <guid isPermaLink="false">https://medium.com/p/a38b70894969</guid>
            <category><![CDATA[ai-overview]]></category>
            <category><![CDATA[seo]]></category>
            <category><![CDATA[indexing]]></category>
            <category><![CDATA[google-search-console]]></category>
            <category><![CDATA[search]]></category>
            <dc:creator><![CDATA[Mikhail Drozdov]]></dc:creator>
            <pubDate>Tue, 01 Sep 2026 08:07:03 GMT</pubDate>
            <atom:updated>2026-09-01T08:07:03.824Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*EnwLEZ0mm2ZibTNv3ICWSQ.png" /></figure><p>What Google’s June 2026 Search Console docs actually list — and what they leave dark</p><p>On 3 June 2026 Google launched dedicated Generative AI performance reports in Search Console. The product is useful. It is also easy to over-read as a query-level AI citation index.</p><p>The Search Central launch post and the Help Center page for the Search report document the same narrow surface: impressions; pages; countries; devices (Search); and dates. The surfaces named for the Search report are AI Overviews and AI Mode. That generative-AI visibility still sits inside the overall Performance Web bag, with a separate dedicated view on top. Help Center answer 16984139 now notes that, as of 31 August 2026, these insights have been rolled out to websites worldwide — so a missing report is less often a “still in subset” story than an impressions-threshold or Exclude-control story.</p><p>What those same pages do not list for this report is the set many site owners will look for first: queries; clicks; CTR; position; or a split series that separates Overview from Mode. Naming the capabilities is not the same as exposing a query dimension. On the published docs, the report answers which URLs showed how often, where, and on which device. It does not, on those pages, answer which queries fired the fan-out or what the AI-feature CTR was.</p><p>Eligibility stays indexing-first. Google’s AI features documentation still requires a page to be indexed and snippet-eligible, with no special AI markup. On top of that stack, Settings → Search generative AI must be Include (the default). Exclude removes links and grounding in those features; Help Center answer 16908024 does not describe it as a ranking signal for the rest of Search, and it does not replace Google-Extended or noindex.</p><p>Date the reporting surface. Do not invent a query index Google has not documented. If product fields expand later — the launch post already says additional metrics may arrive — that will be a new documentation fact, not a retroactive rewrite of June 2026.</p><p>Canonical write-up with field inventory and eligibility stack: <a href="https://casinokrisa.com/blog/gsc-gen-ai-performance-report-2026-06">https://casinokrisa.com/blog/gsc-gen-ai-performance-report-2026-06</a></p><p>Primaries: <a href="https://developers.google.com/search/blog/2026/06/gen-ai-performance-reports">Launch post</a> · <a href="https://support.google.com/webmasters/answer/16984139?hl=en">Help Center report</a> · <a href="https://support.google.com/webmasters/answer/16908024?hl=en">Generative AI control</a> · <a href="https://developers.google.com/search/docs/appearance/ai-features">AI features</a></p><p><em>Mikhail Drozdov — AI Search &amp; Indexing Systems Researcher · </em><a href="https://casinokrisa.com/person/mikhail-drozdov"><em>https://casinokrisa.com/person/mikhail-drozdov</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a38b70894969" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Shared crawl capacity showed up in the docs. That is not a ranking launch.]]></title>
            <link>https://medium.com/@casinokrisa/shared-crawl-capacity-showed-up-in-the-docs-that-is-not-a-ranking-launch-1226176bb4fd?source=rss-a72f34cd294c------2</link>
            <guid isPermaLink="false">https://medium.com/p/1226176bb4fd</guid>
            <category><![CDATA[indexing]]></category>
            <category><![CDATA[information-retrieval]]></category>
            <category><![CDATA[google-search]]></category>
            <category><![CDATA[documentation]]></category>
            <category><![CDATA[crawling]]></category>
            <dc:creator><![CDATA[Mikhail Drozdov]]></dc:creator>
            <pubDate>Mon, 24 Aug 2026 19:22:57 GMT</pubDate>
            <atom:updated>2026-08-24T19:22:57.757Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UuHaYA_n9zD5fLme5Yzc5w.png" /></figure><p>Search engineering has a recurring mix-up: a changelog date on a developer page gets treated as the day a crawler or a ranker changed. On 22 July 2026 Google did change a page. The page is <em>Optimize your crawl budget</em>. What Google wrote about that change is documentation language. I compared the live HTML to a pre-edit archive. The text moved. The publisher still did not claim an algorithm.</p><h3>What the publisher said</h3><p>I am not summarizing a blog roundup. The primary objects are the crawling-infrastructure changelog and its RSS feed.</p><p><em>What:</em> polished and clarified <em>Optimize your crawl budget</em> for clarity, terminology consistency, and flow.</p><p><em>Why:</em> easier to understand, and consistent use of terms such as <em>crawl capacity limit</em>.</p><p>RSS guid: 2026-07-22-Polished-and-clarified-the-crawl-budget-guide. pubDate: 22 Jul 2026 00:00:00 +0000. Live footer: Last updated 2026-07-22 UTC. HTTP Last-Modified when I retrieved the live URL on 24 August 2026: Wed, 22 Jul 2026 18:28:48 GMT.</p><p>Those strings do not mention Googlebot scheduling, hostload computation in production, ranking, or index selection.</p><p>Sources: <a href="https://developers.google.com/crawling/docs/changelog">changelog</a>, <a href="https://developers.google.com/crawling/docs/changelog/crawling_docs_updates.rss">RSS</a>, <a href="https://developers.google.com/crawling/docs/crawl-budget">guide</a>.</p><h3>What the HTML actually did</h3><p>The control text is the Wayback capture from 15 July 2026 06:16:11 UTC of the same path. Against the 24 August live page, thirteen aligned units change. Most of them are housekeeping: “In sum” becomes “Summary”; Index Coverage becomes Page Indexing in two places; a finite-resources paragraph is rewritten; the URL-inventory warning is tightened; Googlebot demand is split into its own paragraph.</p><p>Three units are new sentences, absent from the 15 July capture.</p><ol><li>Crawl does not entail index. For Google Search, a crawled page still has to be evaluated, consolidated, and assessed for suitability.</li><li>Default capacity is conservative and shared in time: if demand exists and the site stays healthy, the limit can be raised.</li><li>Demand is per crawler. Capacity is not. “High demand from one crawler can reduce the capacity available for others.”</li></ol><p>One definition rewrite matters for operators who still think “capacity” means “max parallel connections plus delay.” The live page restates crawl capacity limit (hostload) as the time the server spends holding connections open, counting both parallelism and duration.</p><p>One operational list now names TTFB, 5xx, and HTTP 429 as health signals that can lower the limit. Efficient loading now spells out HTTP 304. Appearance in <em>this</em> guide is an expansion of the page. It is not proof that 304 handling began on 22 July 2026.</p><p>Archive: <a href="https://web.archive.org/web/20260715061611/https://developers.google.com/crawling/docs/crawl-budget">Wayback 20260715061611</a>.</p><h3>What this is not</h3><p>July 2026 Search Central documentation updates list other items, including a 29 July note on social and video platform content in Search Console. They do not list this polish. Crawling-infrastructure edits live on the crawling changelog. In December 2025, when this guide moved, Google wrote that functionality had not changed, only location and some wording. That claim is dated 18 December 2025. I am not back-porting it onto July 2026. I am noting that Google already distinguishes docs location from crawler behavior.</p><p><a href="https://developers.google.com/search/updates">Search Central updates</a>.</p><h3>Interpretation</h3><p>In an indexing-first model, retrieval never sees the open web. It sees an admitted subset. Crawl capacity is one of the admission valves. Publishing a clearer valve diagram is not the same as turning the valve. Shared capacity is the sentence I would keep if I had to keep one: AdsBot, Shopping, and Googlebot are not three independent hostload accounts in the public description. They compete. That is now in the guide. It was not in the 15 July capture. It is still not labeled as a 22 July systems rewrite.</p><h3>Limit</h3><p>I have one archive snapshot, not a continuous redline. DevSite HTML is noisy. Page-level timestamps do not prove every sentence was authored that day. I ran no crawl-rate series. A new public sentence is compatible with documenting old behavior, documenting an earlier unpublished change, or documenting a same-day change. These sources cannot tell those apart. Secondary “new algorithm” recaps are out of scope.</p><h3>Takeaway</h3><p>If your incident review starts from a developer changelog, copy the <em>What</em> and <em>Why</em> before you copy the date into a ranking narrative. For 22 July 2026 the <em>What</em> and <em>Why</em> are documentation. The interesting technical content is the newly public shared-capacity sentence and the crawl-versus-index staging. Treat both as documentation until you have a systems measurement.</p><p>Mikhail Drozdov<br>AI Search &amp; Indexing Systems Researcher<br>https://casinokrisa.com/person/mikhail-drozdov</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1226176bb4fd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The GSC Page Indexing report is not the index]]></title>
            <link>https://medium.com/@casinokrisa/the-gsc-page-indexing-report-is-not-the-index-ff28715ed17d?source=rss-a72f34cd294c------2</link>
            <guid isPermaLink="false">https://medium.com/p/ff28715ed17d</guid>
            <dc:creator><![CDATA[Mikhail Drozdov]]></dc:creator>
            <pubDate>Mon, 24 Aug 2026 11:57:02 GMT</pubDate>
            <atom:updated>2026-08-24T12:20:24.155Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SiHJjBTx3q6f-0-MV7fVJQ.png" /></figure><p>When the Pages report in Google Search Console stops moving, it is easy to treat the freeze as a crawl or ranking event. The summer 2026 delays show why that is a category error.</p><p>Confirmed (journalism, not a Google incident bulletin): Search Engine Roundtable and Search Engine Land documented the Page Indexing report stuck on June 11, 2026 for weeks, then jumping, then flattening into multi-day plateaus. Search Engine Land dated a catch-up to July 3, 2026, with the chart then showing data through June 29. Roundtable later logged further stalls (including a July 1 comment from John Mueller, as quoted there: the team was working on it, no ETA). I did not inspect Search Console properties myself for this draft; I am relying on those contemporaneous reports plus Google’s own help text</p><p>Confirmed (Google Help): The Page Indexing report is a property-level picture of which URLs Google has found and whether they are indexed, with a table of reasons they are not. The URL Inspection tool is the per-URL view of the indexed copy, not a live SERP guarantee. Google’s help for URL Inspection states that “URL is on Google” does not guarantee appearance in results, and that the initial inspection is not a live test.</p><p>The useful distinction: a delayed Pages chart is a delay in reporting. Crawl and selection can still be happening. The live URL test, Google says, cannot even detect several Page Indexing states, including “Crawled - currently not indexed” and “Discovered - currently not indexed.” Those states are assigned at indexing time. If your only dashboard is the aggregate chart, you are looking at a pipeline that is allowed to lag behind the systems that actually fetch and select documents.</p><p>That is why URL Inspection stayed usable as a workaround in the 2026 write-ups, and why overlaying “the report is frozen” onto a spam update is unsafe. Secondary coverage says Google stated the delay was unrelated to the June 2026 spam update; I have not independently verified that statement in a primary Google blog post for this draft, so treat it as reported, not as something I measured.</p><p>Interpretation (mine): indexing diagnostics have two clocks. One is the document’s last crawl and canonical choice. The other is when Search Console’s reporting store lands a new daily (or weekly) point. Confusing them produces false incident response: resubmitting sitemaps, rewriting robots.txt, or blaming quality, when the fetch already happened.</p><p>Practice: for a URL you care about, inspect it. Note last crawl, crawl allowed, indexing allowed, and the Google-selected canonical. Use the live test only to see whether the current HTML is fetchable; do not ask it to predict “Crawled, not indexed.” Keep the Pages report for inventory trends after you know its timestamp is fresh.</p><p>Limits: this is not a controlled measurement of how many properties were delayed, nor of Google’s internal pipeline. I did not reproduce the chart.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ff28715ed17d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Why Google Doesn’t Index Everything]]></title>
            <link>https://medium.com/@casinokrisa/why-google-doesnt-index-everything-2bd6155019ef?source=rss-a72f34cd294c------2</link>
            <guid isPermaLink="false">https://medium.com/p/2bd6155019ef</guid>
            <category><![CDATA[seo]]></category>
            <dc:creator><![CDATA[Mikhail Drozdov]]></dc:creator>
            <pubDate>Mon, 06 Jul 2026 08:08:15 GMT</pubDate>
            <atom:updated>2026-07-06T08:08:15.788Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*S3tfQ_qVgYlSTmzM-fSj6A.jpeg" /></figure><p>One of the most persistent misconceptions in SEO is built on a simple assumption: if Googlebot has seen a page, Google will eventually index it.</p><p>In practice, there is no such equivalence. A page can be fully crawlable, Googlebot can visit it repeatedly, the URL can appear in server logs, and Google can remain aware of the document for months without keeping it in the searchable index. From the outside, this often looks like a technical failure. From the perspective of a search engine, it may be a perfectly normal outcome.</p><h3>Crawling and indexing are different processes</h3><p>When Google discovers a URL, a single action does not follow. There is a sequence of separate processes.</p><p>First, the search engine has to learn that the URL exists. Discovery can happen through an internal link, a sitemap, an external link, a redirect, or a previously known site structure. The URL may then enter a crawl queue. Googlebot may fetch the document. After that, the system still has to process the content, evaluate possible duplication, determine a canonical version, extract text and links, and compare the document with other known pages.</p><p>Only then does the more interesting question appear:</p><p><strong>Does this document need to be in the index?</strong></p><p>This is where many site owners start looking for a technical error that may not exist.</p><h3>Crawl does not mean index</h3><p>I tend to look at indexing through this distinction.</p><p><strong>Crawl asks: did Google retrieve the page?</strong></p><p><strong>Index asks: did Google decide to retain it as a search document?</strong></p><p>These are fundamentally different questions.</p><p>You can build a technically clean page. It can return a 200 status code, use a self-referencing canonical, appear in a sitemap, remain accessible to Googlebot, and receive internal links. None of this creates an obligation for Google to index it.</p><p>Technical accessibility is a condition.</p><p>It is not a guarantee.</p><h3>Why a search engine has to choose</h3><p>The web is too large to treat every URL as equally useful.</p><p>New pages, filters, parameters, archives, copies, automatically generated documents, near-duplicate product pages, programmatic SEO pages, and endless URL combinations appear continuously. A search engine therefore has to solve more than a discovery problem.</p><p>It also has to solve a selection problem.</p><p>Imagine a site that generates 50,000 location pages. On each page, the city name changes, a few numbers are different, and several sentences are rewritten.</p><p>From the CMS perspective, there are 50,000 documents.</p><p>From the search engine’s perspective, the question may be different:</p><p><strong>Are there really 50,000 independent documents here?</strong></p><p>The number of URLs stored on a server and the number of documents worth representing separately in search are not the same thing.</p><h3>The most interesting case: Google knows the page but does not index it</h3><p>For me, this is one of the most useful situations to investigate because the easy explanations have already started to disappear.</p><p>You cannot simply say that Google failed to discover the URL. It did.</p><p>You cannot necessarily say that Googlebot cannot access the page. It can.</p><p>Submitting the sitemap again may not change anything because the URL is already known.</p><p>At that point, I usually start looking at several broader properties of the document and the site around it.</p><h3>1. Document independence</h3><p>Does the page have its own reason to exist?</p><p>Not for the site owner. Not for the page generator. Not for the SEO architecture.</p><p>For the search system.</p><p>If you remove the URL, title, and city name, is the document still meaningfully distinguishable from neighboring pages?</p><p>This is an uncomfortable question for many large-scale SEO projects. It is also a useful one.</p><h3>2. Internal connectedness</h3><p>A page can exist in a sitemap while being almost isolated from the actual site.</p><p>Internal structure is one of the ways a search engine can understand the relative position of a document. If almost nothing links to a URL internally, the page sits deep in the architecture, and its main source of existence is an XML file, that is worth investigating.</p><p>A sitemap communicates that a page exists.</p><p>Site architecture helps communicate where it belongs.</p><h3>3. Semantic duplication</h3><p>Duplication does not always mean identical text.</p><p>Two pages can contain different sentences while solving essentially the same task in nearly the same way. This is especially common on large programmatic sites.</p><p>The text is technically unique.</p><p>The documents are functionally interchangeable.</p><p>A standard plagiarism or text uniqueness checker may see no problem at all. A search system comparing documents at scale may operate on a very different level.</p><h3>4. The broader value of a site section</h3><p>I would not analyze every URL as if it existed in complete isolation.</p><p>If a directory contains a large number of weak, repetitive, or low-value documents, it makes sense to study the behavior of the whole section rather than obsess over a single excluded page.</p><p>There is no need to immediately invent a mysterious penalty. It is more useful to think about how a system allocates crawling, processing, and attention across large groups of documents.</p><h3>5. The document changed after first discovery</h3><p>Some pages are discovered too early.</p><p>A URL may enter the site structure while it still contains almost no useful content. Later, the page becomes complete, but the system has already processed an earlier and much weaker version.</p><p>This does not mean the URL is permanently doomed. It simply means that the history and timing of discovery can be worth considering during diagnosis.</p><h3>Why Request Indexing does not solve the fundamental problem</h3><p>SEO has a familiar ritual.</p><p>A page is not indexed. Search Console is opened. Request Indexing is clicked. Some time passes. The action is repeated.</p><p>Sometimes this helps trigger or accelerate another processing attempt. But the button does not answer the central question:</p><p><strong>Why should the search engine retain this document?</strong></p><p>If the underlying issue involves weak document independence, duplication, architecture, or the broader structure of the site, resubmitting the URL does not change the properties of the document.</p><p>It only asks the system to look again.</p><h3>I prefer to think of indexing as a system decision</h3><p>I find it useful to move away from the idea that Google’s index is a warehouse where pages arrive after passing a technical inspection.</p><p>A better mental model is a continuously revised system of documents.</p><p>Google discovers a URL, retrieves content, compares it with other documents, identifies a preferred version, evaluates its place within the site and among already known material, and then makes decisions about storage and potential use.</p><p>Those decisions are not necessarily permanent.</p><p>A page can enter the index, disappear, return, switch canonicalization, become indexable after changes to internal architecture, or remain known to Google without developing stable search visibility.</p><p>This is why I am increasingly reluctant to describe indexing as a binary technical function.</p><h3>Index and visibility are not the same thing either</h3><p>There is another misconception that appears almost as often: if a page is indexed, it should receive impressions.</p><p>Not necessarily.</p><p>Indexing only means that a document can participate in subsequent search processes. Visibility depends on a different set of conditions: query relevance, competition, intent interpretation, document quality, site signals, context, language, geography, and many other factors.</p><p>It is useful to separate at least three states:</p><p><strong>Google discovered the page.</strong></p><p><strong>Google included the page in the index.</strong></p><p><strong>Google shows the page to users.</strong></p><p>Between each of these states is a separate set of system decisions.</p><h3>What I examine when a page is not indexed</h3><p>My process is usually straightforward at the beginning. I first eliminate technical causes: response status, robots directives, canonicalization, accessibility, redirects, and rendering. Then I look at discovery. How did Google learn about the URL? Are there normal internal links? How deep is the page in the site structure?</p><p>After that, the investigation becomes more interesting.</p><p>I compare the document with neighboring pages. I ask whether it serves an independent search task. I check whether its uniqueness is merely textual. I study the URL cluster rather than one isolated page. I look at which documents Google selects and which ones it repeatedly ignores.</p><p>Only then do I start changing the site.</p><p>Sometimes the problem really is a broken canonical.</p><p>Sometimes the canonical is perfect, the sitemap is clean, the server is fast, and Googlebot visits regularly.</p><p>Technical accessibility alone may simply not be enough for stable inclusion in the index.</p><h3>A final thought</h3><p>The longer I work with search systems, the less I see indexing as a binary technical function. I see it more as an observable decision made by a complex system.</p><p>That is why the question <strong>“Why isn’t Google indexing this page?”</strong> is often too narrow.</p><p>A more useful question is:</p><p><strong>What did Google see in this document, what did it compare it with, and why might a separate place in the index have appeared unnecessary?</strong></p><p>That is usually where a serious investigation into indexing begins.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2bd6155019ef" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Indexing Isn’t Visibility: Search Is a Trust Distribution System]]></title>
            <link>https://medium.com/@casinokrisa/google-indexing-explained-b500bf2ea698?source=rss-a72f34cd294c------2</link>
            <guid isPermaLink="false">https://medium.com/p/b500bf2ea698</guid>
            <category><![CDATA[google-search-console]]></category>
            <category><![CDATA[indexing]]></category>
            <category><![CDATA[seo]]></category>
            <category><![CDATA[search-system]]></category>
            <category><![CDATA[knowledge-graph]]></category>
            <dc:creator><![CDATA[Mikhail Drozdov]]></dc:creator>
            <pubDate>Tue, 10 Feb 2026 08:00:08 GMT</pubDate>
            <atom:updated>2026-02-10T08:00:08.897Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*4Ww56RBG1AIe0Uwe6Z328Q.jpeg" /><figcaption>A page can be stored, crawled, even “known” to Google — and still be effectively invisible. That’s not a bug. It’s the point.</figcaption></figure><p>Search used to feel like a scoreboard.</p><p>You published a page, you waited, you measured rank. If you were good (or lucky), you climbed. If you weren’t, you changed the page until it behaved.</p><p>That mental model still leaks into how most people talk about SEO: as if the web is a list and the algorithm is a judge. But modern search doesn’t behave like a judge. It behaves like a risk manager.</p><p>The shift matters, because it explains the most confusing outcome in Google Search Console: “crawled but not indexed,” or its cousin, “indexed but not visible.” People treat those as technical errors. They often aren’t.</p><p>They’re storage decisions.</p><p>They’re trust decisions.</p><p>And they’re not made at the URL level alone.</p><p>Indexing is storage, not attention</p><p>Indexing is not the moment your page “goes live” to the world. Indexing is the moment a system decides your page is worth keeping in its own memory.</p><p>That framing changes everything.</p><p>When you store information, you pay costs: crawling, rendering, deduplicating, canonical selection, classification, refresh. On a large index, refresh is the real bill. Anything that isn’t predictably valuable becomes expensive inventory.</p><p>So the core question is not “is the page correct?” but “is the page worth repeatedly paying for?”</p><p>People hate this, because it doesn’t feel fair. A page can be true, well-written, and still not be worth storing. A page can be original and still not be worth refreshing. Value is relative, not moral.</p><p>If you want a clean mental model, think of Google as a warehouse.</p><p>Warehouses don’t store every product. They store products that match their distribution plan.</p><p>Visibility is selection + distribution</p><p>Even if a page is stored, it may never be used.</p><p>This is the part most “SEO education” fails to name: being in the index is not the same as being chosen as an outcome. Visibility is an allocation problem. Selection is downstream of storage.</p><p>In practice:</p><p>- Indexing answers: “Should we keep this in memory?”<br>- Retrieval answers: “Should we consider this for this class of queries?”<br>- Selection answers: “Should we actually show this to users, on this surface, right now?”</p><p>You can pass the first gate and still fail the second and third. That’s why “indexed but no traffic” is not a paradox; it’s a typical state.</p><p>The new domain reality: probation, sampling, and low regret</p><p>If you’re building on a new domain, the system’s default posture is conservative. Not because it hates you, but because it has no history of your behavior.</p><p>Search engines don’t just evaluate pages. They evaluate sources over time.</p><p>Consistency, repeatability, and low-regret outcomes are the currency. When a system cannot predict whether your future pages will be coherent, it samples you cautiously. It stores less. It refreshes less. It distributes less.</p><p>This is why people experience “brief spikes” followed by silence and call it a “sandbox.” The useful interpretation is simpler: the system is running a low-regret trial.</p><p>It’s not punishing. It’s calibrating.</p><p>Why “crawled but not indexed” is often a verdict, not a checklist</p><p>Most debugging guides reduce “crawled but not indexed” to a list of possible issues. Lists are comforting; they imply you can fix a single item and unlock the result.</p><p>Sometimes you can. Often you can’t, because the decision is not strictly technical.</p><p>The storage verdict usually comes from one of three forces:</p><p>1) Cost<br>If your pages are heavy to render, slow to fetch, or messy to canonicalize, the system learns that keeping you updated is expensive.</p><p>2) Value<br>If your pages look similar to what the system already has, or they repeat patterns the system has learned to treat as low-signal, your incremental value approaches zero.</p><p>3) Risk<br>If your site behaves unpredictably (shifting canonicals, inconsistent internal linking, unstable taxonomy, sudden topic pivots, noisy URL variants), the system treats you as a risky inventory item.</p><p>Notice what’s missing: “keywords.”</p><p>Keywords still matter as an interface language. They don’t govern storage. Storage is governed by economics and trust.</p><p>The only sustainable strategy: build a source, not a page</p><p>If search is a trust distribution system, then your job isn’t to “optimize pages.” Your job is to become an interpretable source.</p><p>Interpretability is not branding fluff. It’s the opposite: stable claims, stable definitions, stable internal structure, and stable external identity.</p><p>A source becomes interpretable when:</p><p>You explain the same core model repeatedly, in slightly different contexts, without contradicting yourself.</p><p>You make it easy for a machine to understand what you are “about,” and for a human to understand what you stand for, without needing a sales pitch.</p><p>You don’t chase every trend. You build a spine.</p><p>This is why I treat indexing-first SEO as a research problem rather than a tactics problem. The question is not “how do I rank this page,” but “how does a system decide I’m a trustworthy inventory item?”</p><p>Evidence beats self-description</p><p>There’s one more uncomfortable truth: self-claims are weak. Systems have learned to discount them.</p><p>If you say who you are, you’re marketing. If other people say who you are, you’re a source.</p><p>This is the practical purpose of an “evidence hub”: not to brag, but to make identity legible. A place where independent profiles, citations, and references can be audited without hunting across the web.</p><p>This isn’t about gaming Knowledge Panels. It’s about removing ambiguity.</p><p>What this changes in how you publish</p><p>If you accept the model, you stop writing “SEO articles” and start writing a coherent body of work.</p><p>You stop treating each post as a lottery ticket and start treating it as a node in a graph.</p><p>You invest in:</p><p>Clean canonicals.</p><p>Cluster structure (pillars + supporting essays).</p><p>Internal linking that reflects your actual model, not your tag soup.</p><p>And the slow work: accumulating independent evidence that you exist as a stable entity, not just a domain with posts.</p><p>When that foundation is in place, traffic is no longer the fragile outcome of a single page. It becomes a byproduct of being a reliable source inside a system that rewards low-regret distribution.</p><p>That’s the real promise of indexing-first thinking: fewer superstitions, fewer checklists, more clarity about what the system is actually doing.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b500bf2ea698" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Indexing Comes First. Visibility Comes Later.]]></title>
            <link>https://medium.com/@casinokrisa/indexing-comes-first-visibility-comes-later-026504656551?source=rss-a72f34cd294c------2</link>
            <guid isPermaLink="false">https://medium.com/p/026504656551</guid>
            <category><![CDATA[seo]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[information-architecture]]></category>
            <category><![CDATA[search-engines]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Mikhail Drozdov]]></dc:creator>
            <pubDate>Mon, 02 Feb 2026 14:53:09 GMT</pubDate>
            <atom:updated>2026-02-02T14:53:09.399Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YlWBexk_v1qDqjHL887MLA.jpeg" /></figure><p>Search is still often explained as a simple pipeline.</p><p>Crawl.<br>Index.<br>Rank.<br>Traffic.</p><p>That model no longer reflects how modern search systems operate.</p><p>Indexing is not a path to visibility.<br>It is permission to be evaluated.</p><p>A page can be indexed, retrievable, and technically sound, and still be functionally invisible.<br>That outcome is not a failure.<br>It is the system behaving exactly as designed.</p><p>Indexing means the system has decided your content is parseable, classifiable, and worth storing.</p><p>It does not mean the system is willing to expose it repeatedly.</p><p>Visibility requires a different level of confidence.<br>It means the system believes that showing you again tomorrow will not create regret.</p><p>Most indexed pages never reach that threshold.</p><p>Visibility is no longer a relevance decision.</p><p>It is a risk decision.</p><p>Every result shown introduces potential cost: user confusion, corrective reranking, retraining pressure, and trust leakage. Modern search systems optimize for outcome stability under uncertainty, not for theoretical relevance.</p><p>This is why visibility often looks inconsistent from the outside.</p><p>Pages appear briefly.<br>They spike.<br>They fade.<br>They disappear without explanation.</p><p>These are not bugs or penalties.<br>They are controlled exposure tests.</p><p>What actually happens looks closer to an indexing-first model.</p><p>A page is indexed, then surfaced under narrow conditions. Its outcomes are observed. Behavioral signals are evaluated for coherence. Trust is accumulated only if results behave consistently across contexts. Visibility scales only after the system is confident reuse will not require constant correction.</p><p>Most pages stall early in this process.</p><p>They are not rejected.<br>They are simply not trusted enough to repeat.</p><p>This is where traditional SEO assumptions break down.</p><p>Non-ranking is not necessarily misconfiguration.<br>Indexing delays are rarely crawl problems.<br>Visibility loss is often intentional.</p><p>Search engines do not miss good pages.<br>They avoid outcomes they cannot reliably predict.</p><p>The real divide today is between technical certainty and outcome certainty.</p><p>Technical certainty ensures a page can be crawled, indexed, and understood. Outcome certainty determines whether the system feels safe exposing it at scale.</p><p>A page can have clean architecture, perfect markup, and accurate intent matching and still fail visibility. The system is no longer asking whether a page is relevant. It is asking whether the result will behave predictably once deployed repeatedly.</p><p>Indexing grants access.<br>Outcome certainty grants visibility.</p><p>AI-mediated search introduces an additional constraint.</p><p>Many results are now surfaced before explicit questions fully form. This raises the bar dramatically. Only sources that maintain internal consistency, minimize contradiction risk, and exhibit stable topical behavior are safe to surface pre-emptively.</p><p>Indexing alone will never qualify content for this layer.</p><p>When a page ranks briefly and never returns, the system has not penalized it.</p><p>It has rehearsed it.</p><p>The outcome was tested.<br>The signals were noisy.<br>The exposure was revoked.</p><p>No alert is issued.<br>None is required.</p><p>Indexing-first thinking does not mean forcing crawls, inflating coverage reports, or optimizing for inclusion metrics.</p><p>It means designing content that behaves consistently, aligns with system-level intent, and produces outcomes the system can safely repeat.</p><p>The wrong question is how to rank a page.</p><p>The right question is whether the system would feel confident showing it again tomorrow.</p><p>Search is no longer a system of answers.</p><p>It is a system of trust allocation.</p><p>Indexing is access.<br>Visibility is privilege.<br>Trust is the limiting resource.</p><p>And most visibility problems are not technical.</p><p>They are confidence problems, on the system’s side.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=026504656551" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>