Why AI Visibility Scores Differ Between Tools
AI visibility scores differ because tools choose different engines, prompts, denominators, source roles, evidence floors, and freshness windows. Measured on one Machine Relations Index release, engine roster moves the number most: at most 67.5% of cited domains appear in exactly one engine's answers.
AI visibility scores differ between tools because each platform chooses its own engines, prompt set, denominator, source-role rules, evidence floor, and freshness window. Those six choices are usually listed as reasons. They are rarely measured. This update measures all six on a single public release of the Machine Relations Index, so a buyer comparing two scorecards can see which disagreement is large and which is rounding. (Machine Relations Index public data)
Key takeaways
- Engine roster is the first-order term. In the September 19, 2026 release, at most 14,985 of 22,213 cited domains — 67.5% — appear in exactly one of six engines' answers, and the two most similar engines share only 23.3% of their combined domain set. That share is an upper bound: two of the six engines had collection gaps inside this window, and missing days push domains down the agreement ladder. See the correction of September 25, 2026 below.
- The prompt basket is second and larger per brand than most teams expect: the same domain reads 33.96% in its strongest published segment and 0.98% in its weakest on the same release, a 34.7x spread.
- The denominator moves one number without any new data: reddit.com reads 12.67% against all observed answer runs and 33.96% against its best segment's runs, from the same release.
- The evidence floor decides how many numbers exist at all. A citation rate is computable for all 22,213 domains; a confidence grade stands behind 503 of them, so a tool with no floor can publish a rate for roughly 44 times as many domains.
- One choice moved nothing here: competition ranking and a rate-then-runs re-sort produced the same top of every published segment across 880 rows. Not every methodology difference is material, which is why magnitudes beat lists.
- The September 19 Machine Relations Index v2 release reports 125,115 source events, 15,883 observed answer runs, 22,213 domains, and 88 published strata across six engines — four collected continuously, and two, Google AI Overviews and Google AI Mode, with collection gaps the release manifest did not report.
How far each choice moves the number
The six choices are not equal, and a buyer comparing scorecards should spend attention in this order. Every figure below is read from one release — mri_score_v2.0+2026-09-19+0cad03121f60, window May 10 to September 19, 2026, 126 days observed — so the spreads are method spreads, not time spreads. (MRI release manifest)
| Rank | Method choice | Measured spread on one release |
|---|---|---|
| 1 | Engine roster | A single-engine panel sees between 4,888 and 9,522 of 22,213 domains across the three continuously collected engines, a range of at least 1.95x once the three floored engines are included; at most 67.5% of cited domains appear in exactly one engine; best pairwise overlap between two of those four is 23.3% of their combined set |
| 2 | Prompt basket | The same domain reads 33.96% and 0.98% in two published segments of the same release, 34.7x apart |
| 3 | Denominator | 12.67% against all observed runs versus 33.96% against segment runs, same domain, same release |
| 4 | Evidence floor | Rate computable for 22,213 domains, confidence grade behind 503 — a 44x difference in how many numbers can be published |
| 5 | Source-role definition | Ranking classes by the leader's share of its own class inverts the order produced by top-ten share, because two classes hold 10 or fewer domains |
| 6 | Freshness window | Observed runs per date range from 6 to 206, median 47, across 70 dates; a 7-day read quotes a few hundred runs against 15,883 |
| — | Ranking method | No movement on this release: 880 rows across 88 published segments, no change in any top ten |
Engine coverage changes every AI visibility score
A scorecard that measures six engines is not comparable to a scorecard that measures one model or one search surface, and the gap is larger than the word "coverage" suggests. The September 19 release names six engines — ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews and Perplexity — and its manifest reported all six healthy. That manifest field was wrong for three of them, which is the subject of the correction immediately below. A release manifest is not evidence of coverage. (MRI release manifest)
Correction, updated September 26, 2026
An earlier version of this section ranked the six engines against each other by how many distinct domains each cited, and reported the single-engine share and its July-to-September movement as measurements. Those rankings and that movement claim are withdrawn. The Index's own collection recorded zero citations from Google AI Overviews on any monitored query between August 15 and September 24, 2026 — 41 of the last 132 days, including every one of the 35 days between August 15 and this release — and zero from Google AI Mode between September 14 and September 24. The cause was a defect in the Index's Google collection path, fixed September 25, 2026: the prior collector graded a response as a success whether or not it carried content. A distinct-domain pool only accumulates, so missing collection days can remove domains from an engine's set but never add them. Google AI Overviews' 2,279 and Google AI Mode's 7,441 are therefore floors, not measurements, and neither engine's position in a list on this page is a ranking. The same arithmetic runs the other way on the agreement table: missing days push domains down the ladder, so 67.5% at exactly one engine is an upper bound and 1.6% at all six is a floor. Gemini is the third gapped engine, and this note said it was clean until today. The Index recorded zero Gemini citations on 22 intermittent days inside the window — May 11 to 15, then scattered single days from August 11 to September 16. The shape is why it went unnoticed through nine earlier reviews: single scattered days rather than one continuous outage. Gemini's cause is not the Google collection defect and is not yet established. So every Gemini figure on this page is a floor too, and no engine on this page is stated as reaching more domains than Gemini. ChatGPT, Claude and Perplexity collected throughout and their figures are exact; one Claude date was flagged by the same scan and carries exactly one answer run, a scan artifact rather than a gap.
Counted across all 22,213 cited domains, agreement between engines is rare:
| Engines citing the domain | Domains | Share of cited universe |
|---|---|---|
| 1 | 14,985 | 67.5% |
| 2 | 3,516 | 15.8% |
| 3 | 1,792 | 8.1% |
| 4 | 965 | 4.3% |
| 5 | 608 | 2.7% |
| 6 | 347 | 1.6% |
The 1-engine row is an upper bound and the 6-engine row a floor, for the reason in the correction above; the four middle rows can move in either direction. Both Google surfaces are inside the panel these shares are counted over.
Across the three engines collected continuously through this window, reach differs by roughly a factor of two: Perplexity cited 9,522 domains, Claude 5,154 and ChatGPT 4,888 — a 1.95x range, and all three counts are exact. The other three engines are floors, depressed by the collection gaps named in the correction above, and are reported rather than ranked: Gemini at least 7,091, Google AI Mode at least 7,441, Google AI Overviews at least 2,279. Gemini's floor sits above Claude's and ChatGPT's exact counts, so it reaches more domains than both and the margin can only widen; whether it reaches more than Perplexity's 9,522 this release cannot say, and the earlier version of this paragraph, which ranked Perplexity above it, is withdrawn. Read across all six the range is therefore at least 1.95x and may be wider. Overlap between any two engines is thinner still. The most similar pair among the three exactly measured engines, Claude and Perplexity, share 23.3% of their combined domain set; ChatGPT and Claude share 1,223 domains of 8,819 combined, 13.9%. Any pairwise overlap involving Gemini or either Google surface is a lower bound, because removing domains from one set can only lower what it shares with every other set: the 8.6% this section previously reported for ChatGPT and Google AI Overviews is a floor, not a measured overlap.
The concrete case is the most-cited domain in the index. reddit.com ranks #1 of 22,213 domains at 12.67%, cited in 2,012 of 15,883 observed answer runs, and its observed engine list holds four names: Gemini, Google AI Mode, Google AI Overviews, and Perplexity. In this window it was cited in zero ChatGPT runs and zero Claude runs. (reddit.com domain profile) A tool whose panel is ChatGPT-first and a tool whose panel is Perplexity-first will disagree about the single most-cited source on the open web, and both will be reporting what they measured.
Whether the single-engine share is drifting is not readable on these two releases. Measured on July 18, 2026 across 14,680 domains, the single-engine share was 66.4% and full six-engine consensus covered 237 domains; on September 19 the same two fields read 67.5% and 347. The July release sits before the Google collection gap and the September one sits inside it, and a gap inflates the single-engine share, so the 1.1-point rise cannot be separated from the defect. The cross-engine study carries both columns with the Google rows marked as floors for the same reason. (Cross-engine citation agreement, Machine Relations Research) The per-domain map of which engines cite which top sources is tracked separately. (Which AI engines cite Reddit, Paralax)
For buyers, this means an AI visibility score should always disclose the engine roster before anything else on the method sheet. A ChatGPT-only score can be directionally useful, but it cannot be treated as the answer for Google AI Mode, Perplexity, Claude, Gemini, or AI Overviews.
Google's own guidance for generative AI features keeps the work tied to retrievable, reliable public content rather than a separate magic optimization layer. Google says site owners should keep following Search fundamentals for helpful, reliable, people-first content while AI features may use indexed web sources. (Google Search Central)
Prompt baskets change AI visibility denominators
Prompt design determines what the score can mean, and for a single brand it moves the number further than the engine roster does. A brand-defense prompt such as "is Brand X trustworthy" answers a different business question from a buyer prompt such as "best vendor for enterprise data governance" or a source-selection prompt such as "which publications do AI engines cite for B2B software buying questions."
The Machine Relations Index groups observations into strata: a subject category paired with a question type. The September 19 public release reports 157 total strata, 151 collectable, 88 published, 63 collecting, and 49 empty. Those fields prevent readers from treating every topic as equally mature. (Machine Relations Index public data)
Within one release, the same domain's rate depends almost entirely on which strata the prompt basket touches:
| Domain | Strongest published segment | Weakest published segment | Spread |
|---|---|---|---|
| reddit.com | 33.96%, 36 of 106 runs, #1 of 263 (education and training, best tools) | 0.98%, 1 of 102 runs, #153 of 170 (fintech, how buyers choose) | 34.7x |
| forbes.com | 39.39%, 52 of 132 runs, #1 of 236 (consumer finance, best tools) | 0.64%, 4 of 623 runs, #272 of 1,340 (cybersecurity, news topic) | 61.5x |
| medium.com | 46.75%, 36 of 77 runs, #1 of 204 (AI infrastructure, best tools) | 0.76%, #209 of 262 (consumer finance, is it worth it) | 61.5x |
A vendor that reports one blended score across brand, category, competitor, and news prompts may be useful for an executive dashboard. It is weaker for diagnosis unless the underlying prompt basket is visible, because a brand can move a blended score by changing its prompt mix without changing anything a machine reads.
The denominator moves the number without new data
Two rates computed from identical observations can differ by a factor of three, because they divide by different run counts. An overall citation rate divides by every observed answer run in the release; a segment rate divides by the runs observed in that category and question shape.
reddit.com reads 12.67% overall — 2,012 cited runs of 15,883 observed — and 33.96% in education and training, best tools, where the denominator is 106 runs. Nothing about the source changed between those two numbers. (reddit.com domain profile)
This is the disagreement most often mistaken for a data error. When two tools quote different numbers for the same brand in the same week, the first question is not "who is wrong" but "what is on the bottom of each fraction."
Evidence floors change whether thin signal becomes a score
A stricter evidence floor produces fewer but more interpretable scores, and the difference in volume is large. MRI v2 publishes a source-segment citation rate only after a segment clears at least 10 observations across at least seven distinct run dates. Below that line, the segment is marked collecting rather than scored. (Machine Relations Index public data)
On the September 19 release, a citation rate is computable for all 22,213 cited domains, while a confidence grade stands behind 503 of them: 14 at A, 58 at B, 431 at C. The remaining 21,710 domains — 97.7% — are marked collecting. At the segment level, 88 of 157 strata are published.
That is the measured cost of the floor, and it runs in one direction: a tool that publishes a number for every domain it ever observed can show roughly 44 times as many scored domains as a tool with this floor, using the same underlying observations. Completeness can be cosmetic if the score is based on one run, one model, one geography, or one volatile news day.
Statistical guidance outside AI visibility points in the same direction. NIST's handbook treats proportion intervals as a method attached to assumptions and sampling conditions, not as a decorative number after the fact. AAPOR's survey best-practice guidance similarly emphasizes design, data collection, analysis, and transparency before readers judge a result by size alone. (NIST) (AAPOR)
Source roles change AI visibility rankings
A source-citation score changes when the tool separates owned pages, earned media, communities, social platforms, and reference sites — and the ranking flips depending on how the class leader's share is expressed. Two brands can have the same mention count while relying on very different evidence layers. One may be cited through its own blog. Another may be cited through reviews, media coverage, documentation, community threads, analyst pages, or partner sites.
Measured on the September 19 release, the size of the class decides the shape of the number:
| Source class | Domains in class | Leader | Leader's share of class | Top-ten share of class |
|---|---|---|---|---|
| Search or media platform | 10 | youtube.com | 94.6% | 100% |
| Wire and press-release distribution | 9 | businesswire.com | 42.4% | 100% |
| Community and social | 27 | reddit.com | 46.6% | 97.3% |
| Editorial publication | 1,224 | medium.com | 6.0% | 25.7% |
| Vendor-owned | 823 | microsoft.com | 2.7% | 18.8% |
| Other observed source | 18,655 | healthline.com | 0.4% | 1.8% |
Rank the classes by concentration using top-ten share and the small classes look identical at 100%. Rank them by the leader's share of its own class and the order changes, because a top-ten share is meaningless in a class holding nine or ten domains. A vendor that reports "share of voice by source type" is making this choice whether or not the method sheet says so.
That distinction matters because AI systems synthesize from multiple sources. The original Generative Engine Optimization paper formalized generative engines as systems that gather and summarize information from multiple sources, creating a visibility problem for website and content creators. (GEO paper, arXiv)
For operators, source role is diagnostic. If a brand's owned pages appear but independent sources do not, the correction layer is not just more owned content. It may require earned authority, third-party corroboration, clearer documentation, or better citation architecture.
Freshness windows change AI visibility volatility
A score measured today can disagree with a score measured last week because the sample behind "last week" is small. The September 19 MRI manifest identifies a public artifact generated on September 19, with data observed from May 10 through September 19 and 126 days observed. That release identity keeps the number tied to a specific window. (MRI release manifest)
The observed run cadence explains most short-window volatility. Across 70 observed dates in this release, runs per date range from 6 to 206 with a median of 47. A seven-day window therefore rests on a few hundred runs, while the full window rests on 15,883. Two tools quoting "the last week" and "the last quarter" are not quoting the same sample size, and the shorter one will move more for reasons that have nothing to do with the brand.
The buyer question should be simple: can the vendor reproduce the same report with the same date, engine roster, prompt set, and denominator?
The choice that did not matter
Not every methodology difference is material, and saying so is part of the method. Rank assignment is a real choice — competition ranking, dense ranking, or a re-sort by rate and then by run count all produce defensible orders when two domains tie.
On this release it changed nothing worth reporting. Across 88 published segments and 880 ranked rows, competition ranking and a rate-then-runs re-sort agreed at the top of every segment, with no ties inside any top ten. An earlier release showed a small divergence, so this is a release-dependent property rather than a constant — which is the point. A method difference is worth arguing about only after someone measures how far it moves the number.
A practical checklist for comparing AI visibility tools
CMOs should compare AI visibility tools by method before comparing scorecards by number, in order of measured impact. The score is the output. The method is the product.
| Method field | What to ask | Measured effect on the number |
|---|---|---|
| Engine roster | Which models and answer surfaces are measured? | Largest: at most 67.5% of cited domains appear in exactly one engine; per-engine reach varies by at least 1.95x across the three continuously collected engines |
| Prompt basket | Are prompts brand, category, competitor, or buyer-intent prompts? | Up to 61.5x between segments for one domain on one release |
| Denominator | Is the score based on answer runs, citations, mentions, positions, or traffic? | 2.7x on identical data for the most-cited domain |
| Evidence floor | How many runs and dates are required before a score is published? | Decides whether 503 or 22,213 domains carry a number |
| Source roles | Are owned, earned, community, reference, and social sources separated? | Inverts class rankings when classes differ in size by three orders of magnitude |
| Freshness window | What dates does the report cover, and can the result be reproduced? | A 7-day read rests on a few hundred runs, the full window on 15,883 |
| Ranking method | How are ties and equal rates ordered? | No effect on this release across 880 ranked rows |
| Missingness | Are failed runs, empty answers, and collecting segments disclosed? | 21,710 of 22,213 domains sit below the grade line |
This checklist does not require every vendor to use the Machine Relations Index method. It requires every vendor to disclose enough method for a buyer to interpret the output.
What brands should do after the score changes
The right response to a changed AI visibility score is diagnosis before correction. If the score fell because a key answer engine stopped citing an earned-media source, the response is different from a decline caused by stale owned pages or a prompt-basket change.
Use a three-layer operating model:
| Layer | Question | Output |
|---|---|---|
| Detection | Are we mentioned, cited, omitted, or misdescribed? | Scorecards, prompt logs, citation lists, sentiment flags |
| Diagnosis | Which engine, prompt, source role, and date changed? | Root-cause map by answer surface and source layer |
| Correction | What public evidence should change? | Updated pages, stronger third-party proof, clearer claims, schema, earned media |
Measurement is the instrument, not the work. AI visibility tools are measurement infrastructure, and a dashboard cannot make a brand more citable by itself.
This is where Machine Relations becomes a useful operating frame. Machine Relations is the discipline of earning AI citations and recommendations for a brand by making that brand legible, retrievable, and credible inside AI-driven discovery. It treats measurement as one layer of a larger source architecture, not as a substitute for the evidence machines need to cite.
FAQ
Why do AI visibility tools show different scores for the same brand?
AI visibility tools show different scores because they often measure different engines, prompts, source roles, dates, and denominators. The engine roster is the largest single term: on the September 19, 2026 Machine Relations Index release, at most 67.5% of 22,213 cited domains appear in exactly one of six engines' answers, so two tools with different panels are reading largely different source universes. That figure is an upper bound — two of the six engines had collection gaps inside the window, per the correction above.
Which methodology choice moves an AI visibility score the most?
Engine roster moves it most across a whole universe of sources, and prompt basket moves it most for a single brand. Measured on one release, a single-engine panel sees between 4,888 and 9,522 domains across the three engines collected continuously through the window, and at least that wide a spread once the three floored engines are counted, while one domain's rate ranges from 33.96% to 0.98% depending on which category and question shape the prompts land in.
Is one AI visibility score the true number?
No single AI visibility score is the true number unless the business question and method are specified. A citation-rate score answers how often sources were cited in observed answer runs; a brand-mention score answers whether the brand appeared in a prompt sample.
What is a good evidence floor for AI visibility measurement?
A good evidence floor discloses how many observations and run dates stand behind the score. MRI v2 uses at least 10 observations across at least seven distinct run dates before publishing a source-segment citation rate; on the September 19 release that leaves 503 of 22,213 domains carrying a confidence grade and the rest marked collecting.
Should brands buy an AI visibility tool before fixing source architecture?
Brands should use measurement to guide source architecture, not replace it. A tool can detect omissions, citations, and sentiment, but the correction work usually requires clearer owned pages, better third-party corroboration, consistent entity signals, and more extractable proof.
How should CMOs evaluate an AI visibility vendor?
CMOs should ask for the engine roster, prompt sample, denominator, evidence floor, source-role taxonomy, release date, missing-data policy, and reproducibility rules. If the vendor cannot explain those fields, the headline score is not ready for budget decisions.
Sources and method
Every figure in this article is read from one public release of the Machine Relations Index, mri_score_v2.0+2026-09-19+0cad03121f60, sha256 0cad0312..., generated September 19, 2026, window May 10 to September 19, 2026, 126 days observed, 15,883 observed answer runs across six engines, two of which — Google AI Overviews and Google AI Mode — had collection gaps inside the window that the release manifest did not report, per the correction above. Engine breadth, segment rates, class shares, and run cadence were computed from that artifact; the reddit.com figures were re-read at the live domain profile on September 19, 2026. Ranks are quoted with the population they were computed against. The July 18, 2026 comparison figures are quoted from the earlier Machine Relations research page named below.
- Machine Relations Research. "Machine Relations Index public data." September 19, 2026. https://machinerelations.ai/data/machine-relations-index.json
- Machine Relations Research. "MRI release manifest." September 19, 2026. https://machinerelations.ai/data/mri-release-manifest.json
- Machine Relations Index. "Source domain profile: reddit.com." September 19, 2026. https://machinerelations.ai/index/domains/reddit.com
- Machine Relations Research. "Cross-Engine Citation Agreement: Which Sources All AI Engines Cite." July 18, 2026. https://machinerelations.ai/research/cross-engine-citation-agreement-source-consensus-2026
- Paralax. "Which AI Engines Cite Reddit? Four of Six, and the Split Has Held for Two Months." September 18, 2026. https://paralax.ai/blog/which-ai-engines-cite-reddit-per-engine-source-map-2026
- Google Search Central. "Guide to Optimizing for Generative AI Features on Google Search." https://developers.google.com/search/docs/fundamentals/ai-optimization-guide
- Aggarwal et al. "GEO: Generative Engine Optimization." arXiv. https://arxiv.org/abs/2311.09735
- NIST/SEMATECH. "7.2.4.1. Confidence intervals." e-Handbook of Statistical Methods. https://www.itl.nist.gov/div898/handbook/prc/section2/prc241.htm
- AAPOR. "Best Practices for Survey Research." https://aapor.org/standards-and-ethics/best-practices/