Discussion starter: an AI search question, posted with an answer from Niraj Raut to open the thread. If you have dealt with this on a site, reply with what you saw, especially where it differs.
- AI Search
Ahrefs found AI Overview citations from top-10 pages fell from 76% to 38%, with a method change. Does your own tracking show ranking and citation decoupling?
Disclosure: Niraj Raut, who posted this answer, runs the SEO consultancy linked at the end of it.
1 reply
-
adminAdmin
The short answer is that the public evidence supports partial decoupling between classic rankings and AI Overview citations, but the 76% to 38% change should not be read as a trend line. Ahrefs itself says its detection improved between the two studies, so part of the drop is measurement. The more useful finding is structural: a large share of cited pages do not rank well for the query being tracked, which is what you would expect if citations are tied to rankings for related sub-queries rather than to the visible query alone.
So when your own tracking appears to show “ranking and citation decoupling”, the first question is what you are measuring rank against. If you only track position for the head query, you will see decoupling almost by construction. If you also track the sub-questions the answer covers, the relationship often looks less like decoupling and more like the coupling moving to a different set of queries. Public evidence on that second point is still thin, so treat it as a hypothesis worth measuring, not a settled fact.
What the Ahrefs numbers support, and what they do not
Search Engine Journal’s coverage of the Ahrefs study gives the detail that matters for interpretation.
Finding What it supports What it does not show 38% of cited pages ranked in the top 10 for the same query (March 2026, about 863,000 keywords), versus 76% in July 2025 Top-10 ranking for the tracked query is far from a requirement for citation A clean halving over time. Ahrefs says its parsing now detects more citations, so the samples are not directly comparable About 31% of all cited URLs ranked beyond position 100 for that query (36% in the organic-only view) Many citations come from pages with no visible ranking for the tracked keyword That those pages rank for nothing related. Rank was measured only against the original query BrightEdge reported about 17% top-10 overlap with a different method and dataset The direction is consistent across vendors A precise figure. Two vendors, two methods, two very different numbers Timing overlaps a Gemini 3 upgrade in January 2026 and Google’s documented query fan-out Plausible reasons for citation behaviour to change Which cause explains how much of the change I would treat the 38% figure as a fair snapshot of the current state and the comparison with 76% as mostly unusable. Whether the underlying relationship weakened, and by how much, cannot be answered from these two studies.
The mechanism: citations are chosen for the sub-questions
Google’s AI features documentation says AI Overviews and AI Mode may use a “query fan-out” technique, “issuing multiple related searches across subtopics and data sources” to develop a response, and that this lets them show “a wider and more diverse set of helpful links”. The same page sets a floor: a page must be indexed and eligible to be shown in Google Search with a snippet, with no additional requirements.
An analogy helps here. Think of an analyst asked to brief a manager on “best CRM for a small agency”. They do not run one search and cite the top results. They run separate searches for pricing, integrations, client portals and migration effort, then cite whichever sources answered each sub-question best. The footnotes in the final brief ranked well for those sub-searches, not necessarily for the title of the brief.
The technical version runs in stages that are worth keeping separate:
User query triggers an AI Overview↓Fan-out: related sub-queries are issued (not visible to you)↓Retrieval: candidate pages come from the index for each sub-query↓Citation selection: which retrieved passages support the synthesised answer?↓Answer synthesis with supporting linksRank trackers measure the first query. Citations are decided after the fan-out and retrieval stages. The gap between those two is where most of the apparent decoupling lives. None of this says classic ranking signals stopped mattering: retrieval still comes from Google’s index, and the eligibility floor is the same as for a normal snippet.
Hypothetical example:
A SaaS comparison page ranks 14th for “best project management software for agencies”, so a tracker marks it as outside the top 10. The AI Overview for that query covers client portals, time tracking and pricing. The same page ranks 2nd for “project management software with client portal”. It gets cited. In a head-query-only report, this looks like decoupling. In a sub-query report, it is a page ranking well for the part of the answer it was cited for.
Why your own tracking can exaggerate or hide the effect
Before concluding anything from in-house data, I would rule out these artefacts:
- Citation volatility. The same query can return a different set of cited sources on repeated checks. A single daily snapshot overstates how stable a citation is, and a single miss overstates loss. Sample each query several times across a week.
- Rank depth. Since Google stopped supporting the
&num=100parameter in September 2025, many tools changed how deep and how often they check. “Not in the top 100” means different things in different tools. - Parsing scope. Tools differ in whether they capture citations hidden behind “show more”, inline links and side panel links. Ahrefs’ own method change shows how much parsing alone can move the numbers.
- URL versus domain. The cited URL is often a different page on your domain from the one that ranks. Measure at both levels, or you will report decoupling that is really internal page choice.
- Organic versus SERP features. Ahrefs’ organic-only view differed from the all-results view, which suggests some cited pages rank through other features rather than classic listings.
- Location and device. AI Overview presence and sources vary by market. Many features launch in the US first, so a UK or Australian panel may behave differently.
How to set up tracking that answers the question
If you want in-house evidence rather than vendor averages, build it around rank buckets and sub-queries.
- Build a query panel. A few hundred queries that trigger AI Overviews in your market, segmented by intent (informational, comparison, transactional) and by topic cluster.
- Capture citations repeatedly. For each query, record cited URLs and domains from several runs over a week. Use citation frequency (for example, cited in 4 of 6 runs) rather than a yes/no flag.
- Record head-query rank to a fixed depth. Use the same depth for every query and every run, and note which ranking URL you own.
- Derive and track sub-queries. For each head query, list 3 to 5 likely sub-questions from the AI Overview’s own section structure, People Also Ask and follow-up prompts. Track your best-ranking URL for each.
- Compare citation rate by rank bucket. Calculate citation frequency for head-query rank buckets (1 to 3, 4 to 10, 11 to 20, 21 to 100, not found) and separately for best sub-query rank buckets.
Pattern in your data Likely interpretation What I would do Citation rate rises steeply with head-query rank Still coupled for your topics Keep prioritising rankings on the tracked queries Weak link to head-query rank, strong link to sub-query rank Coupling has moved to fan-out queries Map and strengthen coverage of the sub-questions, often on the same page or tightly linked pages Weak link to both Selection is driven by something else, or your sub-query guesses are wrong Inspect cited competitors: source type, freshness, directness of the answer passage Cited, but a different URL from the ranking one Internal page choice rather than decoupling Check whether the cited page answers the sub-question more directly, and whether consolidation makes sense What Google and Bing reporting can add
Search Console’s generative AI performance reports reached all eligible sites on 31 August 2026. They show impressions only, with no clicks, CTR or queries, filterable by page, country, device and date. That means you cannot join AI impressions to rank per query. You can join at page level: compare pages with meaningful AI impressions against their average Web position and see whether low-ranking pages still collect AI exposure. Remember that AI Mode is already counted inside the Web search type totals, so do not treat Web and generative AI figures as separate pools.
Bing Webmaster Tools’ AI Performance report is useful for a different reason: it shows a sample of grounding queries behind citations. They are Bing’s, not Google’s, but they are the closest first-party view available of the queries used at the retrieval stage. Comparing them with the keywords you track is a quick way to see how far apart the two sets are.
What this should change in practice
Decoupling at the head-query level does not justify a separate “citation optimisation” workstream built on formulas. Google’s May 2026 guide to generative AI features says foundational SEO still applies and that no special files, chunking or AI-specific markup are needed. What it does change is the unit of keyword research: the sub-questions behind a topic deserve tracking in their own right, because that is plausibly where citation eligibility is won.
It also changes reporting. A citation is exposure, not a visit. Clicks from AI Mode reach GA4 as ordinary Google organic traffic according to practitioner reports, so AI Overview and AI Mode value will not show up in a separate channel. Report citation frequency, AI impressions and organic clicks as distinct metrics rather than blending them.
The decision rule I would use
Do not reallocate budget away from rankings on the strength of the 76% to 38% headline. Run the rank-bucket comparison above on your own query panel for four to six weeks. If citation frequency tracks sub-query rank much more closely than head-query rank, shift keyword research and content planning towards the sub-questions. If it tracks neither, study the cited competitors before changing strategy. If it still tracks head-query rank, your topics have not decoupled, whatever the industry averages say.
Need help with this on your own site?
Niraj Raut works with businesses in Nepal, Australia, the UK, Europe and the US on technical SEO, ecommerce SEO, local SEO and AI search.
Add a reply Cancel reply
Replies are open to approved members. Log in or apply to join.