AI Overview Sources Changed for 52% of Keywords in New Study

Google changed at least one AI Overview source for 52% of tested keywords when the exact same search was repeated, according to a study published on August 13, 2026 by Position Digital. In the same dataset, only 33.2% of cited URLs also ranked in the organic top 10, which means the citation layer and the blue-link layer are not returning the same answer.

The dataset is small and hand-collected: 857 citations, 321 unique URLs, 31 US SaaS keywords, 92 repeated runs, one account, one country, one session per day. That is enough evidence to change how you measure AI visibility. It is not enough to justify rewriting a site.

What Is AI Overview Citation Volatility?

AI Overview citation volatility is the tendency of Google to return a different set of source links for the same query on repeated, identical searches. In this study it appeared as at least one changed citation for 52% of keywords, with only 48% of keywords returning the same source set every single time.

The volatility was not evenly spread across the citation list. The first cited URL stayed identical across every run for 86% of keywords, while consistency dropped for positions further down the sidebar. That pattern is the useful part: the top slot behaves like a preference, and the remaining slots behave like a rotating pool of acceptable sources.

For anyone reporting AI visibility to a client or a manager, this single finding invalidates the most common measurement method in the industry: opening a browser once, screenshotting an AI Overview, and calling it a result.

What the August 13 Study Actually Measured

The study collected 857 total citations covering 321 unique URLs across 227 root domains, gathered manually from Google AI Overviews for 31 US SaaS keywords across seven niches. Each keyword was run three to four times, producing 92 runs used to measure how often citations change.

The headline numbers worth carrying forward:

Read that last row carefully. A 64.7% grounding rate sounds like a damning accuracy statistic until you notice the denominator is 17 claims. Six claims flipping would move that number by more than 30 points.

The Limitations Are the Most Important Part of This Study

The study is observational, single-account, single-country, and small, and the authors say so directly. Nothing in it was a controlled experiment, and nothing in it is guidance from Google.

The constraints that matter before you quote any of these numbers:

• One account, one US session per day. No logged-out control, no personalization test, no cross-locale comparison. Personalization and location effects are not separated from model behaviour.
• 31 keywords in specific B2B SaaS niches. Resource management, data collaboration, salary benchmarking, AI visibility tracking, ecommerce shipping, enterprise AI. Nothing here covers news, geopolitics, developer documentation, or the informational queries that most of our traffic comes from.
• Three to four runs per keyword. With that many runs, a keyword that rotates one source occasionally and a keyword that rotates constantly can produce the same "changed at least once" flag.
• Correlations are not causes. A domain with high traffic and high citation counts may share an unmeasured cause with both.
• Observational, not Google guidance. Google has published nothing that confirms a source pool, a rotation policy, or a stability preference for the top citation.

I would still act on this study, but only in one direction: it tells you your measurement is too noisy, not what your content should say.

Why Only 33.2% of Cited URLs Rank in the Top 10

Two thirds of cited URLs came from outside the organic top 10, which is the strongest available signal that AI Overview source selection runs on a different retrieval and selection path than the classic ranking list. Ranking still helped: pages at positions one to three were cited far more often than pages at seven to ten.

The practical readthrough is that "get to page one, then get cited" is an incomplete plan. A page can be selected as a supporting source for a passage it answers precisely while never ranking for the head term, and a page can own position one while being ignored by the answer that sits above it.

This is also why a single tracking rank column cannot represent AI visibility. Rank is one number per keyword. Citation presence is a probability per keyword, per run, per position. Our running Google algorithm updates hub tracks the ranking side, and the citation side needs its own dataset next to it.

How Reliable Are AI Overview Claims About Your Content?

Only 64.7% of the 17 specific claims checked in the study were directly supported by a source cited in the same passage, with the remainder contradicted, unsupported, inherited from a source error, or grounded elsewhere in the answer. That is a small sample, but the failure modes are the interesting output, not the percentage.

Two of those failure modes should worry publishers more than hallucination does. The first is inheritance: when Google repeats a factual error that already exists on a cited page, the error travels with your brand attached. The second is conflict resolution: when sources disagreed, the vendor number won in at least one case.

If your page carries a number that is stale, ambiguous, or unlabelled, an answer engine can propagate it into a summary that millions of people read without ever visiting you. Precision on your own pages is a distribution decision now, not a copy-editing preference.

Our Analysis: Citation Visibility Is a Distribution, Not a Position

The right mental model after this study is that AI Overview citation is a probability distribution over a pool of eligible sources, and you should measure it the way you measure a flaky test rather than the way you measure a rank.

That reframing changes three things about how a small site like abhs.in should work.

First, sample size beats screenshots. If 52% of keywords change at least one source between identical searches, then a single check has a coin-flip chance of misreporting your presence in the rotating positions. Any claim of "we got cited" needs a run count attached.

Second, the first position is a different objective from the rest. An 86% stability rate for the top citation means displacing it is a slow, content-quality problem, while entering the rotating pool below it is a faster, coverage-and-specificity problem. Those deserve different work.

Third, topical fit outperforms brand weight. With 88.4% of vendor citations coming from topically relevant vendors, depth inside a defined subject is the affordable lever for a site without a large authority profile. That matches what already works here: narrow, specific, numbers-first posts inside a cluster.

One thing this study explicitly cannot support: any claim that adding schema, FAQ blocks, or llms.txt guarantees citations. Nothing was tested against a control. Structure makes an answer easier to extract; it does not buy a slot.

A Reproducible AI Overview Citation Monitoring Protocol

The protocol below turns the volatility finding into a repeatable measurement you can defend, and it is deliberately boring: fixed keyword set, fixed run count, fixed recording format, no spot checks.

Run it like this:

1. Freeze a keyword set of 20 to 30 queries per cluster and never edit it mid-quarter.
2. Run each query five times in a clean session, logged out, with the same location setting, and record every citation in display order.
3. Record run timestamp, position of each citation, whether your brand was named in the prose, and whether an AI Overview appeared at all.
4. Repeat weekly on the same weekday to keep index-refresh effects roughly constant.
5. Report presence rates with the run count in the same sentence. "Cited in 4 of 5 runs" is a result; "we are cited" is not.
6. Treat any week-over-week move smaller than one run out of five as noise.
7. Log claim-level errors separately, because a wrong number attributed to you is a bigger problem than a missing citation.

If you want to automate the answer-checking step with a model, price it before you commit, because repeated runs multiply token cost quickly. Our LLM API pricing tracker is the fastest way to model that per-run cost.

What This Changes for abhs.in GEO

For this site, the study confirms the strategy and changes the reporting, not the writing. Definition-first H2s, specific numbers, named entities, and tight FAQ blocks stay because they make passages extractable, but no post here should claim they guarantee an AI Overview citation.

Three concrete adjustments follow. Measure citation presence across five runs per query instead of one, split reporting between the stable first slot and the rotating tail, and audit the exact numbers in high-impression posts because answer engines can repeat an error faster than we can correct it. The wider strategy for this stack, including Bing and answer-engine surfaces that feed ChatGPT browse, sits in our AI search playbook, and the traffic-shape argument behind it is in the zero-click AI Overviews playbook.

The honest summary of the state of AI visibility measurement in August 2026 is that the industry is publishing point estimates from single sessions, and one small manual study just showed how unstable those point estimates are.

Sources

• Position Digital: How Does AI Overview Pick its Sources? An Analysis of 300+ SaaS Citations, August 13, 2026
• Google Search Central: AI features and your website
• Google Search Console documentation

Key Takeaways

• 52% of keywords returned at least one changed AI Overview source across 92 repeated runs, so single spot checks are unreliable.
• 86% of keywords kept the same first citation, meaning the top slot is stable while lower slots rotate.
• 33.2% of cited URLs ranked in the organic top 10, confirming citation selection is not the ranking list.
• 88.4% of vendor citations were topically relevant, which favours narrow topical depth over broad domain authority.
• 64.7% of 17 checked claims were directly supported by a cited page, a small sample that still exposes real error-inheritance risk.
• For developers: measure citation presence as a rate across at least five runs per query, logged out, same location, and report the run count with every number.
• What to watch: whether larger, multi-locale, logged-out replications reproduce the 52% volatility rate outside US B2B SaaS keywords.