A visibility score is only honest about what it holds constant.
Four AI engines answered the same 15 commercial prompts on June 6, 2026. In all 60 top-five comparisons the audit reported, they shared no exact URL. That limits what one score can contain.
A business reading a single AI visibility number may be reading two different kinds of quantity added together. One is a property of its own pages. The other is what a particular engine did with a particular question on a particular day. Only the first has a fixed value, and a single number cannot show which of the two moved.
An AI visibility score can hold constant only what a business publishes
The only inputs to an AI visibility score that hold still between two readings are the business's own published pages. Some of the questions it can score: does a page state what the business does, is the business described the same way everywhere, can the content be crawled and parsed, are reviews and location present in the markup? Those are properties of the pages, and they do not depend on which engine is asked.
The AIOInsights evaluator retrieves the public content a crawler would and applies a versioned rubric. The same fetched content under the same rubric version returns the same score. The determinism covers the scoring, not the fetch: review markup updates on its own, JavaScript can render differently, and a CDN can serve two requests different bytes. When a score moves, the cause is in the fetched content or in the rubric version, and both can be named.
Which URLs an AI engine cites is a different kind of quantity. It belongs to an engine, a question and a day, and it has no fixed value to score. Put the two in one number and the stable part lends its credibility to the unstable part.
Four AI engines given the same prompt almost never cite the same URL
In all 60 top-five comparisons it reports, four AI engines given the same commercial prompt shared no exact URL. That is the starkest result in an audit posted to arXiv on September 19, 2026 by Benjamin Tannenbaum. The audit ran 15 fixed commercial prompts through ChatGPT, Microsoft Copilot, Google and Perplexity and recorded 589 citation observations on June 6, 2026, covering 528 unique URLs and 356 domains. Of those URLs, 96.4% appeared in only one engine.
The audit's limits apply to every figure this essay takes from it: one author, 15 English-language prompts about AI search visibility software, one day of cross-engine data and one day pair for change over time. The author founded Aiso Boost Ltd., which develops software for measuring visibility in AI search, and the data was collected through its infrastructure.
In 84.9% of the same-prompt engine pairs the audit could compare, the two engines shared no cited URL at all. Mean pairwise similarity (Jaccard) was 0.0079 and the median was zero. Only ten prompts were observed on all four engines, so the pair counts fall short of 15 prompts times six engine pairs.
On those ten prompts, a chance baseline that draws lists of the same lengths at random from each prompt's pooled URLs predicts a similarity of 0.1272. The observed figure was 0.0072, or 5.7% of it. The baseline rules out one explanation: that the engines diverge only because they cite lists of different lengths. It cannot say why they diverge. Separate indexes, different ranking and different rewriting of the query all fit the data. The measured result does not need the reason. What one engine cites says very little about what another will.
AI engines churn their citations overnight, each at its own rate
Between June 5 and June 6, 2026, with the prompts held fixed, turnover in cited URL sets was 82.6% for ChatGPT, 76.6% for Microsoft Copilot and 45.5% for Perplexity, in the Tannenbaum arXiv audit of AI engine citations. The mean across the 41 same-prompt pairs was 67.0%. Google had no June 5 observations and is not in the comparison. The author treats these figures as a diagnostic, not an estimate of daily volatility, because the change may reflect the collection system as well as the engines.
The per-engine split carries more than the mean. Every engine measured churned, and the fastest turned over nearly twice the share of the slowest. A score that folds in citations from all three would move overnight by an amount set by which engine it leans on.
Repeated askings within one collection show the same instability at a smaller scale. The AIOInsights live sample, re-collected weekly, asks Claude with web search enabled four local-service questions eight times each: best estate planning attorney in Thousand Oaks California, best family dentist in Pasadena California, best fee-only financial advisor in Austin Texas, and best personal injury lawyer in Phoenix Arizona. In the collection dated November 14, 2026, 35% of the businesses named across those 32 answers appeared only once.
A business that asks one engine one question once cannot tell whether it sits in the core that repeats or the tail that appears once and is gone. Being cited once and being absent once are both weak evidence. The informative quantity is how often, over many askings.
One engine-agnostic citation score turns a weighting choice into a number
The same AI engine citation observations produce different single numbers depending on how the engines are weighted. In the Tannenbaum arXiv audit, the share of each prompt's pooled URLs that one engine surfaced averaged 42.6% for ChatGPT, 24.3% for Google, 23.6% for Perplexity and 11.4% for Microsoft Copilot, across the ten prompts all four engines answered.
Count the four engines equally and the blend is 25.5%. Weight them by Similarweb's May 2026 share of worldwide web visits, as Search Engine Journal reported it (ChatGPT 53.9%, Gemini 27.9%, Perplexity and Copilot 1.3% each, with Gemini standing in for Google), and the blend is 35.8%. Nothing new was observed. More than ten points came from a decision about weights.
Neither figure describes an engine. On those ten prompts ChatGPT cited 18.8 URLs per prompt and Copilot 4.7, so a page competes for far fewer slots on one than on the other. A business absent from Copilot and a business absent from ChatGPT face different work, and the blended figure points to neither.
The remedy is also an average, a frequency over many askings, and the difference matters. Repeating one question on one engine samples one process. The frequency estimates a property of that question and engine pair, and the estimate tightens as askings accumulate. Averaging across engines combines four processes, and the result depends on weights someone chose. More data makes the first number more accurate. It only makes the second more stable.
A blended AI citation score stays a weighting choice even beside its breakdown
The strongest case for a blended AI citation score is that it can be published with a per-engine breakdown and an interval, so nothing is hidden. A reader who wants the detail has it, and a reader who wants one figure has that too. As a labeled summary printed beside its breakdown, that is a fair design, and the argument against it is narrower than a ban.
The weights are a choice, not a measurement. The engine shares in the Tannenbaum arXiv audit blend to 25.5% when every engine counts equally and to 35.8% when engines count by web traffic. The breakdown is the measurement. The blend adds an editorial decision and no information.
The interval misleads. An interval around a blend treats the gap between engines as spread around a central value. In the Tannenbaum arXiv audit, with a median pairwise overlap of zero, the gap between engines is the finding.
The blend also breaks a score it is folded into. A deterministic score returns the same number for the same content. Add a citation component and the same pages return a different total next week, and a reader cannot tell whether the site changed or an engine did. A blended figure can sit beside the breakdown as a summary. It cannot sit inside a number that claims same input, same output.
AI citation divergence shrinks at the domain level and does not disappear
AI citation divergence shrinks when engines are compared by site instead of by exact URL, and it does not disappear. Counting exact URLs overstates disagreement: two engines citing different pages on the same site register as zero overlap, and a business cares whether its brand appears, not which page. BrightEdge data reported by Search Engine Journal in April 2026 put overlap in cited sources between pairs of AI search surfaces at 16% at the low end and 59% at the high end. The URL figures also come from one small audit, and a larger design could shrink them.
Both points reduce the size of the claim. Neither reverses it. The 59% high was between Google AI Mode and Google AI Overviews, two surfaces Google operates. Gemini, also Google's, overlapped with AI Overviews at 34% and with AI Mode at 27%. Three surfaces from one company spread from 27% to 59%, and the full range ran down to 16%. That leaves no common value for a blend to estimate. The argument does not need turnover to be two thirds. It needs engines to differ by enough that one figure hides the difference, and the site-level range shows that they do.
The objection does change the method. A tracked series should record brand and domain mentions alongside exact URLs, because URL churn overstates how unstable a brand's presence is.
A deterministic readability score measures page fit, not citation
A deterministic readability score measures whether a business's pages can be read and understood by a machine, and no evidence yet ties that score to how often AI engines cite the business. AIOInsights has not shown that a higher score raises citation frequency on any engine, and the evidence in this essay cuts against assuming it. If engines cite almost entirely different URLs for the same prompt, no single property of a page can explain selection on all of them.
The Tannenbaum arXiv audit draws the same boundary. It says its results do not invalidate page scoring done without a live engine. They identify what such a score estimates: page quality, or the fit between a page and a query. End-to-end visibility also depends on whether a given engine exposes the page and whether it then selects it. The paper argues for reporting page fit, observed exposure, conditional selection and final visibility as distinct quantities. An earlier arXiv study of 602 prompts across ChatGPT, Google and Perplexity adds that how many sources an engine cites, and how much each cited page shapes the answer, diverge by platform.
The AIOInsights score claims the first of those four quantities and nothing further. The evaluator does not query any engine's recommendation system and says so on its methodology. A low score is a finding about the pages. A high score is not a forecast of citation.
AI citations belong in a tracked series, one engine per line
The unit of AI citation measurement is a fixed question asked of a named engine, repeated on a schedule, with each result stored and dated. The reading is a frequency over time for that question and engine pair, never a single yes or no.
Fix the wording of each question, because a reworded question starts a new series. Report each engine on its own line. Record brand and domain mentions as well as exact URLs. Record a miss as a measured zero and an unasked day as unknown, because the two are different findings.
Print the readability score next to the series and never inside it. Side by side, the two lines can test the unproven relationship directly: whether pages that score higher are cited more often on a given engine. Inside one number, that test cannot be run.
Being cited once and being absent once are both weak evidence.
A larger audit could overturn this. If many prompts, many days and several authors showed the engines converging on the same sources and holding them, a citation figure would have one value to estimate and could be scored like a page. The published evidence so far is the single audit described above, and it points the other way. Until that changes, an honest report carries two lines and labels them: what the pages are, and what each engine did.
Further reading
- Benjamin Tannenbaum, Scoring With the Engine: Retrieval Exposure, Cross-Engine Divergence, and the Limits of Engine-Agnostic GEO Scores. The audit of four engines on 15 commercial prompts that supplies the cross-engine overlap and day-over-day turnover figures. arxiv.org/abs/2609.22655
- Zhang Kai, He Xinyue and Yao Jingang, From Citation Selection to Citation Absorption. A 602-prompt study showing citation behavior diverges by platform, used as independent support that engines are not copies of one process. arxiv.org/abs/2604.25707
- Search Engine Journal (Roger Montti), report on BrightEdge data comparing cited sources across AI search surfaces. Used in the objection section for site-level overlap between engine pairs. searchenginejournal.com/comparison-of-ai-citation-patterns-offers-strategic-seo-insights
- AIOInsights, How AI Recommends Businesses: A Live Share-of-Voice Sample. Our own repeated sample of one engine across four local-service questions, the position this essay extends. aioinsights.com/insights/how-ai-recommends-businesses
- AIOInsights, evaluator methodology on the homepage. The statement that the check is deterministic and what it does not measure. aioinsights.com
- Search Engine Journal (Greg Jarboe), ChatGPT, Gemini & Claude Lead AI Visibility, Is It Time To Stop Tracking Perplexity? Reports Similarweb's May 2026 share of worldwide web visits by AI assistant, used as the market-share weights for the blended citation figure. searchenginejournal.com/chatgpt-gemini-claude-lead-ai-visibility-is-it-time-to-stop-tracking-perplexity