Your AI Mention Share Is Really Three Numbers: What Engine-Level Splits Reveal That a Blended Score Hides

Suppose your Generative Engine Optimization (GEO) dashboard says you hold 45.8% of AI mentions in your category, against 27.5% for your nearest rival. That looks settled. Now split the same number by engine: on ChatGPT you are second, at 32.3%, behind the rival at 38.5%. On Gemini you are first with 57.3%. Same brands, same month, same prompts. One number became three, and two of the three tell a different story than the headline.

This is not a rounding artefact. It is the normal condition of AI visibility measurement in 2026, and it matters more than most of what gets argued about in GEO. A blended mention share is an average over systems that read different corpora, retrieve differently, and format answers at different lengths. Averaging them produces a tidy number that no single user experience corresponds to.

The Data

The figures below come from daily automated querying of ChatGPT, Gemini and Claude in Japanese at 06:00 JST, with brand-free buying questions ("which X should I use?") and a count of which tracked brands appear. Shares are share of mentions across the tracked brand set in each category, not an absolute probability of appearing in any given answer. The published window runs 9 September to 10 October 2026, and the detail pages disclose the collection cadence: ChatGPT and Gemini are currently retrieved every other day, Claude roughly daily through a separate route. That resolves to 16 valid observations per category — four days each from ChatGPT and Gemini, eight from Claude. Days that fail collection are excluded from the averages rather than counted as zeroes.

Hold onto those per-engine day counts. They turn out to be the most important numbers on the page.

Three Categories, Three Shapes of Disagreement

Engineer recruiting: the leader depends on the engine

Brand	Blended	ChatGPT (4 days)	Claude (8 days)	Gemini (4 days)	Spread
Levertech	45.8%	32.3%	47.8%	57.3%	25.0 pts
Geekly	27.5%	38.5%	21.8%	22.2%	16.7 pts
Workport	26.7%	29.2%	30.4%	20.4%	10.0 pts

Three brands, three engines, and the ranking is different in each. Levertech runs from a weak second on ChatGPT to a dominant first on Gemini — a 25-point swing in the same brand's share, from the same prompts, over the same window. Geekly is first on ChatGPT, second on Gemini, third on Claude. The published engineer-recruiting category data shows the per-engine rows directly, which is the only reason the blended 45.8% is legible as an average rather than a fact.

If you are Geekly, the blended number says you are losing by 18 points and the ChatGPT row says you are winning. Both are true statements about different systems. Which one belongs in a board deck depends entirely on where your buyers actually ask.

Hotel booking: one brand is second on one engine and last on another

Brand	Blended	ChatGPT	Claude	Gemini	Spread
Booking.com	23.8%	26.9%	22.2%	22.3%	4.7 pts
agoda	18.4%	21.0%	15.5%	18.9%	5.5 pts
Expedia	18.2%	26.8%	17.0%	10.8%	16.0 pts
Rakuten Travel	14.8%	9.8%	17.0%	17.6%	7.8 pts
Jalan	12.7%	9.0%	14.1%	14.9%	5.9 pts
Ikyu	12.2%	6.5%	14.3%	15.6%	9.1 pts

Expedia is the case worth studying. On ChatGPT it is effectively tied for first at 26.8%. On Gemini it is last of six at 10.8%. The blended 18.2% places it a hair behind agoda and reads as a solid mid-pack position, which is the one thing it demonstrably is not on either engine.

There is a second pattern in the same table that has nothing to do with any individual brand. Add up the top three on each engine: 74.7% on ChatGPT, 56.2% on Claude, 58.8% on Gemini. ChatGPT's answers in this category are markedly top-heavy — a short recommendation list concentrates share and truncates the tail, which is why the three Japanese domestic brands all post their worst numbers there and their best on Gemini. That is a difference in answer format, not in brand reputation, and no amount of content work changes it. The hotel booking category series carries the per-engine rows behind these figures.

Cloud accounting: the control case

Brand	Blended	ChatGPT	Claude	Gemini	Spread
freee	38.2%	35.3%	43.8%	35.4%	8.5 pts
Money Forward	31.3%	32.3%	26.3%	35.4%	9.1 pts
Yayoi Accounting	30.5%	32.4%	30.0%	29.1%	3.3 pts

Here the engines broadly agree. freee leads on all three, though Gemini puts Money Forward level with it at 35.4%, and no brand moves more than about nine points across engines. This is what a category looks like when the three systems have converged on the same answer set — and it is the case where a blended number is a reasonable summary rather than a blur. Note that this window published seven valid Claude days rather than eight, for 15 observations.

The useful comparison is between categories, not within them. Engineer recruiting has a maximum spread of 25 points; cloud accounting, 9.1. One of those categories can be reported as a single number. The other cannot.

The Blended Number Weights Each Engine Equally — Check the Arithmetic

This is worth doing by hand once, because it changes how much the headline figure deserves.

Levertech: (32.3 + 47.8 + 57.3) / 3 = 45.8, which is exactly the published blended figure. Booking.com: (26.9 + 22.2 + 22.3) / 3 = 23.8, again exact. freee: (35.3 + 43.8 + 35.4) / 3 = 38.2. The blend is an unweighted mean of the three per-engine shares.

That is a defensible choice, and it has two consequences nobody puts on a dashboard:

• A third of your headline number rests on four observations. ChatGPT contributes one third of the blend from four valid days, as does Gemini; Claude's eight days also count one third. The engine with twice the sampling gets no extra weight, so the noisiest inputs are the most influential per observation. With four draws, a ten-point engine gap is not reliably distinguishable from sampling variation.
• Equal weighting is not neutral. It implicitly asserts that your buyers use the three assistants in equal proportion. Almost nobody's audience does. A usage-weighted blend — even a crude one built from your own analytics referrer data — would produce a different leader in both of the divergent categories above.

Neither point is a flaw specific to one vendor. Every blended AI visibility score has a weighting scheme inside it, and most products do not tell you what it is. Recomputing the mean from the per-engine rows is the fastest way to find out.

Why Engines Disagree

Four mechanisms, in rough order of how much they explain:

• Answer shape. An engine that returns three options and an engine that returns six distribute share very differently, regardless of what either "thinks" of your brand. The hotel table above is mostly this effect. It is the dominant source of divergence in categories with more than four or five credible players.
• Retrieval versus recall. An answer assembled from live web results reflects what currently ranks and gets cited. An answer generated from parametric memory reflects a training corpus with a cutoff. The same question routed through those two paths yields different brand sets, and the routing is not always visible to the user, let alone to a tracker.
• Corpus composition. Models trained with different mixes of licensed data, forum text and regional web content inherit different priors about which brands are the obvious ones. This is the part that looks most like reputation and moves the slowest.
• Non-determinism and small samples. Ask any of these models the same buying question twice and the brand list can change. Only the aggregate over many draws is stable, and four draws is not many. Some of every engine spread in this article is noise.

The practical test for separating the last mechanism from the first three: does the gap persist in the same direction across two consecutive windows? Expedia's ChatGPT-versus-Gemini split is large enough and structurally explainable enough to act on. A six-point engine gap in a three-brand category is not, and should be left alone until a second window agrees.

A Reporting Protocol for Per-Engine Data

1. Publish the per-engine rows next to the blend, always. Not in an appendix. If the maximum spread in your category exceeds roughly ten points, lead with the engine breakdown and treat the blended figure as context.
2. Choose your priority engine from audience data, not from results. The temptation is to headline whichever engine you are winning on. Pick instead from referrer data, sales-call evidence and the assistant your buyers' employers have actually deployed, then commit to it for the programme.
3. Diagnose low share by its engine signature. Weak on ChatGPT but fine elsewhere, in a crowded category, usually means you are below the list-length cutoff — the fix is being an obvious top-three answer, not publishing more. Weak on every engine means you are missing from the answer set entirely, which is a different and larger project.
4. Require per-engine valid-day counts. A tool that reports a blended score without disclosing how many observations each engine contributed cannot tell you whether a movement is real. Ask for the counts before you ask about features.
5. Never celebrate a blended move you have not traced to an engine. A three-point rise in the blend can be a single engine's four-day read drifting. Trace it to a row, or do not report it.
6. Keep a separate line for AI Overviews. Citation inside Google's AI summaries behaves differently again, and most mention-share trackers do not cover it. The conceptual background — including why AI answers assemble competitor sets around the user's job rather than around industry boundaries — is laid out in this AIO fundamentals explainer.

What This Means for Tool Selection

Once you accept that per-engine data is the unit of analysis, engine coverage stops being a feature-list line and becomes the binding constraint. A tool that does not cover the engine your buyers use cannot answer your question at any price.

As the market stands in October 2026, Profound sits at the enterprise end with the broadest coverage, spanning ten or more surfaces including Perplexity, Copilot and AI Overviews. Peec AI is the mid-market option with strong source attribution. Otterly.ai is the cheapest entry point, with third-party comparisons listing a Lite tier around USD 29 per month — but also noting that Claude and Gemini are paid add-ons rather than base coverage, which is exactly the kind of detail that decides whether two of the three rows in the tables above exist for you at all. In Japan, Mieruca GEO from Faber Company covers ChatGPT and Gemini plus AI Overviews and AI Mode citation status, with consulting attached, at a published entry point of JPY 49,800 per month plus initial fees.

AIOPulse, whose public data this article uses, is the narrow option: ChatGPT, Gemini and Claude only, with per-engine rows and valid-day counts published on every category page. Its site tracking plan is JPY 19,800 per site per month excluding tax, including five keywords, with additional keywords at JPY 5,000 each per month, no initial fee and month-to-month cancellation. There is no free trial — the first month is charged at signup. It does not cover Perplexity, Copilot or AI Overviews citations, so if the AI Overviews surface is central to your case, one of the broader tools is the right call. The vendor publishes its own comparison against Mieruca GEO, which is worth reading with the obvious caveat about who wrote it. Verify all pricing directly before purchase; published rates in this category move quarterly.

For context on scale when comparing public datasets, the AIOPulse public rankings covering 56 categories and 252 brands publish the sample counts and collection-failure handling alongside the percentages, which is what made the arithmetic in this article checkable from outside.

FAQ

Should we optimise for one engine or all three?

Measure all of them, prioritise one. The work that moves a brand into an answer set — being unmistakably associated with a specific job, getting cited by sources these systems trust, having comparison-shaped content that survives summarisation — benefits every engine. But the trade-offs are engine-specific, and a programme trying to improve three divergent numbers simultaneously tends to optimise none of them.

Is a per-engine gap real or just noise?

With four valid days behind an engine's number, treat gaps under about ten points as unresolved and gaps that reverse between windows as noise by default. A gap that is large, persistent across two consecutive windows, and explainable by a structural difference such as answer length is worth acting on.

Which engine should carry the most weight?

Whichever one your buyers use, which you should establish from your own referrer data rather than from market-share headlines. Build a crude weighted blend from those proportions and keep reporting it alongside the equal-weight version; when the two disagree about who leads, that disagreement is itself the finding.

Do these Japanese-language figures apply to English-language markets?

The mechanisms do; the numbers do not. Answer-length effects, retrieval-versus-recall routing and non-determinism are properties of how these systems generate text, not of any one language. Use figures like these to calibrate how wide a per-engine spread can plausibly get before you assume your own dashboard is broken — not to estimate your own position.

Takeaway

A single AI visibility score is a convenience, not a measurement. Behind it sit three systems that disagree — sometimes by 25 points on the same brand, sometimes about who leads the category outright — combined by a weighting scheme most dashboards never state. Split the number by engine, check how many observations each engine contributed, pick the engine your buyers actually use, and report the spread. The blended figure is still worth having; it is just the least informative number on the page.

Disclosure: AIOPulse is developed by Marketing Cloud Lab, which also operates this site. All AIOPulse figures in this article were retrieved on 10 October 2026 from its public category pages, covering the published window of 9 September to 10 October 2026 (9 September to 9 October for cloud accounting), and AIOPulse pricing was read from its published pricing section on the same date. Competitor coverage and pricing for Profound, Peec AI, Otterly.ai and Mieruca GEO reflect publicly available comparisons and vendor pages as of October 2026 and were not independently audited.