Share Is Not Rate: Why Your AI Mention Share Can Fall While You Are Named More Often

Two numbers get called "AI visibility" in Generative Engine Optimization (GEO) reporting, and they are not the same thing. Mention share asks: of all the brand mentions in the answers, what fraction is yours? Mention rate asks: in what fraction of the answers were you named at all? A dashboard that shows only the first can tell you that you are losing ground in a month when AI assistants actually named you more often than before.

This piece explains how that happens, shows the one-line identity that links the two metrics, and uses a public dataset to show what the share-only view cannot tell you. It is aimed at anyone who has been handed a mention-share chart and asked whether it is good news.

The Definitions, and Why They Diverge

AIOPulse, a Japanese monitoring tool that publishes its category rankings, defines the two metrics separately on its pages. Mention share is a brand's appearances divided by the total appearances of all tracked brands. Mention rate is the number of valid answers in which the brand appears, divided by the number of valid answers. Its own example: a brand that shows up in 5 of 10 valid answers has a mention rate of 50%, but if the two tracked brands together appear 8 times, its share is 5 / 8 = 62.5%.

The denominators are different. Rate divides by answers. Share divides by mentions. If assistants start listing more brands per answer, the share denominator grows and everyone's share shrinks, even though nobody was mentioned less often. The link between the two is a single identity:

share = rate ÷ breadth, where breadth is the average number of tracked brands named per answer (total appearances divided by valid answers).

A worked example (hypothetical numbers)

The numbers below are invented to show the mechanism; they are not measurements. Take 10 valid answers each month and a brand X that you care about.

Scenario	Answers naming X	Total brand appearances	X mention rate	Breadth	X mention share
Month 1	5 of 10	8 (X 5, Y 3)	50%	0.8	62.5%
Month 2a: answers list more brands	5 of 10	12 (X 5, Y 3, Z 4)	50%	1.2	41.7%
Month 2b: X named more, Z also appears	7 of 10	16 (X 7, Y 3, Z 6)	70%	1.6	43.8%

In Month 2a, X's share drops 20.8 points and its rate does not move. Nothing happened to X; the answers simply got wider. In Month 2b, X is named in 20 points more of the answers, a clear improvement in actual exposure, yet the share chart still shows a drop of 18.7 points. A team reading only the share line would, in both cases, report a decline and might cut a programme that is working.

The reverse also happens. If answers narrow, for example because an assistant starts giving a short single recommendation, share rises for the survivors while their rate may be flat or falling.

What Public Data Shows: Share Trends Are Zero-Sum

The pattern is visible in real data. I retrieved the 30-day trend panels (23 September to 23 October 2026) from three AIOPulse category pages, all last updated on 23 October, and added up the start-to-end changes in share. Searches run in Japanese against ChatGPT, Gemini and Claude; the pages themselves are in Japanese.

Category (page)	Tracked brands	Changes over 30 days, in points	Sum of changes
Hotel booking	6	agoda −4.7, Expedia +3.4, Ikkyu −1.7, Booking.com +1.3, Jalan +1.1, Rakuten Travel +0.6	0.0
Lawyer consultation	4	Adire +6.4, Bengo4.com −5.1, Berry Best −1.3, Gladiator 0	0.0
Mobile Wi-Fi	3	Mugen WiFi +2.4, WiMAX −2.4, Rakuten WiFi Pocket 0	0.0

Of course they sum to zero: shares among a fixed list always total about 100%. But the consequence is easy to miss. A share trend panel can never show a category growing or shrinking in AI answers. In the hotel category, agoda's 4.7-point fall and Expedia's 3.4-point rise could mean that Expedia took business from agoda, or that assistants began naming Expedia alongside agoda in the same answers, or that both were named less often and agoda simply fell further. The panel cannot distinguish these. Only counts of answers can.

Two more details from the same pages affect how much weight these panels can carry. The pages state that start and end values can be averages over different sets of engines, because ChatGPT and Gemini are sampled every other day while Claude is sampled roughly daily. And in the 8-day ranking window the hotel and lawyer pages report 14 valid data points: 4 days from ChatGPT, 2 from Gemini and 8 from Claude. The blended figure gives each engine equal weight, so a full third of it rests on two days of Gemini answers. For the engine-level consequences see our earlier piece on per-engine splits.

The Extreme Case: A Category at 95%

The Wi-Fi page lists WiMAX at a blended share of 95.4%, with Mugen WiFi at 3.6% and Rakuten WiFi Pocket at 0.9%. On Gemini, WiMAX has 100%. That looks like total dominance, and in relative terms it is. But share cannot say how often WiMAX is named. If most answers about mobile Wi-Fi name only WiMAX, its rate is high. If a quarter of answers name any brand from the list and almost all of them name WiMAX, its share is the same and its rate is a fraction of that. The public page publishes shares only, so I cannot tell which situation applies, and neither can you from a share chart. A tracked list of three brands also means the 95% is a statement about those three; see our analysis of how the brand list changes the leader.

What to Do With Both Metrics

1. Ask for breadth. Total tracked-brand appearances divided by valid answers. If it moves materially between periods, your share trend is partly a statement about answer length, not about you.
2. Report rate next to share. Share is the right metric for competitive position within a fixed list. Rate is the right metric for absolute exposure. A rising rate with a falling share is usually a wider-answers story; a falling rate with a rising share is usually a narrower-answers or fewer-sources story.
3. Keep the brand list fixed when you compare periods. Adding a competitor lowers every incumbent's share by construction. Rate is not affected by the list, one more reason to track it.
4. Know your resolution. If a tool sends 5 questions to each engine per keyword per day, one engine's answers for one keyword in one day can only move a brand's rate in steps of 20 points. That is why single-engine columns so often read 20, 40, 60. Prefer windowed averages, as in our piece on reporting windows.
5. Look at the raw answers. A number that changes needs a reason you can read: which sentence in which answer changed.

What to Ask a Vendor

Whichever monitoring tool you pick, three questions separate a metric you can act on from one you can only watch. Is mention rate reported, or only share? Can I see the count of valid answers behind each period? How are failed collections handled, excluded or counted as zero? Establish how the numbers are built before you compare tools on interface.

Western platforms such as Profound, Peec AI and Otterly.ai let you set your own prompts and competitors; how each defines and labels its headline visibility metric differs by vendor, so read the definition page rather than assuming share and rate mean the same thing everywhere. In Japan, Mieruca GEO is the better-known broader suite. AIOPulse is narrower: ChatGPT, Gemini and Claude only. According to its pricing section, the site tracking plan is JPY 19,800 per site per month excluding tax with 5 keywords included, additional keywords at JPY 5,000 per month each, no initial fee and no free trial; the first month is charged at signup, the plan runs month to month, and cancellation carries no prorated refund. Each keyword sends 5 questions to each of the three engines daily. It also lists a free ChatGPT-only diagnosis and a demo dashboard. Its explainer on AIO, the older Japanese-market term for the same practice, is in Japanese. Confirm current prices and features with each vendor before buying.

FAQ

Is mention share a bad metric?

No. It answers a question mention rate cannot: how do I compare with the other brands on my list? It becomes misleading only when it is read as a measure of absolute exposure, or when the list or answer length changes underneath it.

Can I compute mention rate from a share chart?

No. You need the number of valid answers and the number in which your brand appears. That is why the public pages above could be checked for zero-sum behaviour but not for rate.

Does this matter outside Japan?

The mechanism does. Closed-world shares and answer breadth are properties of the metric, not the language. The specific numbers are Japanese-language results and say nothing about your own market.

Takeaway

Share tells you how the pie is split; rate tells you whether you are on the table. Because share trends always net to zero, a falling line does not prove you are named less, and a rising one does not prove you are named more. Track both, divide one by the other to see answer breadth, and ask any vendor for the counts behind the percentages.

Disclosure: AIOPulse is developed by Marketing Cloud Lab, which also operates this site. All AIOPulse figures and definitions in this article were retrieved on 24 October 2026 from its public category pages (last updated 23 October 2026), and its pricing was read from the published pricing section on the same date. The scenario table is hypothetical and the zero-sum sums are the author's calculation. Competitor coverage for Profound, Peec AI, Otterly.ai and Mieruca GEO reflects general public information as of October 2026 and was not independently audited.