Two Windows, Two Winners: Why Your AI Mention Share Depends on the Reporting Period You Pick

Here is an uncomfortable exercise. Pull your AI visibility numbers for last month, then pull them for last week, and see whether the two reports agree on who is winning. In a category with any real competition, there is a good chance they do not — and the gap is not a bug in the tool. It is what happens when a noisy, low-sample measurement is summarised over two different periods, and it is the most common way Generative Engine Optimization (GEO) reporting goes wrong.

This is not hypothetical. On 10 September 2026 we pulled the same public Japanese-language dataset two ways: the category detail pages, which average a fixed 30-day window, and the summary leaderboard, which reflects a more recent read. Same source, same day, same brands. In three of the categories we checked, the two views named different leaders.

Three Categories Where the Window Decided the Winner

The underlying data comes from daily automated querying of ChatGPT, Gemini and Claude in Japanese at 06:00 JST. Shares are share of mentions across the tracked brand set in each category, not an absolute probability of appearing in any given answer.

Cloud CRM: a 0.9-point lead is not a lead

Brand	30-day average (10 Aug – 9 Sep 2026)	Latest read (10 Sep 2026)
Salesforce	30.7%	—
HubSpot	29.8%	26.7% (1st)
Zoho CRM	27.6%	—
kintone	11.9%	—

Over the month, Salesforce leads by nine tenths of a point over HubSpot, with Zoho a further two points back. On the more recent read, HubSpot is on top. Anyone at HubSpot writing a board slide on 10 September could accurately claim the number one position in Japanese-language AI answers for CRM; anyone at Salesforce could accurately claim the same thing on the same day. Both would be citing the same published CRM category data. Neither would be lying.

The right conclusion is the boring one: at a three-point spread across three brands, this category has no leader. It has a top tier.

Budget SIM carriers: the leaderboard flips

Brand	30-day average
Rakuten Mobile	20.7%
IIJmio	16.2%
mineo	15.2%
UQ Mobile	12.1%
LINEMO	10.9%
ahamo	9.0%
Y!mobile	8.0%
povo	7.8%

Rakuten Mobile leads the 30-day average by 4.5 points. On the latest read, IIJmio holds the top spot at 18.1%. In a flat eight-brand category where the top four sit inside nine points of each other, one period's sampling is enough to reorder the top two. The daily series for this category also carries the detail that makes the flip legible: the tracked month contained 19 data points, and collection failed on four days each for ChatGPT and Gemini, with those days excluded from the averages rather than silently treated as zeroes.

Programming schools: same brand, 20 points apart

The starkest case involves no reordering at all. DMM WEBCAMP shows 47.4% on the 30-day average for 10 August to 9 September, and 27.3% on the latest read. Below it, Tech Camp (16.7%), TechAcademy (16.3%), RUNTEQ (8.6%), Samurai Engineer (6.9%) and Potepan Camp (4.1%) split the remainder over the month.

A 20-point difference in the same brand's number, from the same source, on the same afternoon. If your GEO retainer is judged on "did our mention share go up", a switch of reporting window is more than large enough to manufacture the answer in either direction.

Why the Gap Exists

Three mechanisms, and it is worth being precise about which is doing the work, because only one of them is fixable by better tooling.

• Non-determinism. Ask the same model the same buying question twice and you can get two different brand lists. Only the aggregate distribution over many samples is stable. Any single day is a draw from that distribution, not a reading of it.
• Low sample counts. The CRM and programming-school windows above are built on 16 data points each across three models over 30 days. That is enough to rank a 47% brand against a 4% brand with confidence. It is nowhere near enough to resolve a 0.9-point gap. Small denominators also produce suspiciously round numbers, which is why several categories in the same dataset show a leader at exactly 100% — flight comparison, inheritance law, company incorporation. Those are not monopolies. They are thin brand sets and few observations.
• Real movement. Underneath the noise there is genuine drift. An earlier retrieval of the same CRM category, covering 10 July to 9 August 2026, had Salesforce at 37.0%, HubSpot 31.0%, Zoho CRM 25.5% and kintone 6.4%. Across the two adjacent windows Salesforce fell 6.3 points and kintone rose 5.5. That is a real trend line, and it is only visible because both readings used the same window length.

The first two are properties of the measurement instrument and no vendor can eliminate them. The third is the thing you are actually paying to observe. The entire job of a reporting protocol is to stop the first two from being mistaken for the third.

A Reporting Protocol That Survives Contact With a Board Meeting

1. Fix the window before you look. Write down the period length — 28 or 30 days is a reasonable default — and commit to it for the whole programme. The moment a window is chosen after seeing the data, the report has stopped being a measurement.
2. Compare like windows only. This month's 30-day average against last month's 30-day average. Never a rolling average against a spot read, which is exactly the error the CRM example produces.
3. Report the tier, not the rank. If three brands sit inside three points, say so. "Top tier alongside HubSpot and Zoho" is both more honest and more useful than "second place", which will invert next week.
4. Set a minimum detectable change in advance. With roughly 16 observations per window, treat anything under about five points as unresolved. Publishing that threshold up front is what stops a noise-driven bump from being claimed as a campaign result.
5. Check the denominator behind every figure. How many samples, how many collection failures, and how failures were handled. A missing day dropped from an average and a missing day counted as a zero produce very different charts from identical raw data.
6. Log model releases against your series. A shift that moves every brand in the category at once, in the same week a major model update ships, is an instrument change and should be annotated as one.

Step five is the one teams skip and the one that most often explains an inexplicable chart. If you want to see what a transparent version looks like before buying anything, the public rankings across 56 categories and 252 brands publish sample counts and failure days alongside the percentages, which is enough to sanity-check the numbers in this article yourself.

What to Ask a Vendor

Most GEO tool evaluations turn into a feature-count contest. For the window problem specifically, four questions do the work.

• Can I set the aggregation window myself, and does the default match across every view in the product? A dashboard whose headline tile and detail page use different periods will generate exactly the contradiction above, in-house, every month.
• How many samples per prompt per day? This single number sets your minimum detectable change. A tool that will not answer it is asking you to trust a distribution it has not described.
• How long is raw daily history retained? Retention shorter than your reporting cycle means you can never recompute a past period under a corrected method, and you cannot backfill a comparison you did not think to run at the time.
• How are collection failures represented? Excluded, zeroed or interpolated — all three are defensible, and only one of them should be silent.

On the market as it stands in September 2026, Profound sits at the enterprise end with the broadest engine coverage including Perplexity, Copilot and AI Overviews, from around USD 499 per month. Peec AI is mid-market with strong source attribution, from around USD 95. Otterly.ai is the cheapest entry point at around USD 29 and covers a wide spread of countries and languages. In Japan, Mieruca GEO from Faber Company starts at JPY 49,800 per month plus setup and covers ChatGPT and Gemini, with AI Overviews citation tracking and AI-referred conversion analysis that the cheaper trackers do not attempt. AIOPulse starts at JPY 19,800 per month per site (five keywords included, no free trial) and tracks all three of ChatGPT, Gemini and Claude, but is a mention-share tracker rather than a full-funnel suite. The vendor-published AIOPulse versus Mieruca GEO comparison sets out the split in more detail, including the areas where Mieruca is the stronger product — read with the obvious caveat that one of the two parties wrote it. Verify all pricing directly before purchase; published rates move.

FAQ

Is a 30-day window always right?

It is a good default because it averages out weekday effects and usually contains enough samples to be readable. Shorten it only if your sampling rate is high enough that a shorter period still clears your minimum detectable change, and lengthen it if your category is volatile enough that you cannot otherwise separate trend from noise. What matters far more than the specific length is that it never changes.

Our tool only shows a current snapshot. Is that useless?

Not useless, but it answers a different question. A snapshot tells you roughly whether you are in the answer set at all, which is a genuine binary worth knowing. It cannot tell you whether anything you did worked, because that requires two comparable periods. If the tool does not retain history, export daily and build the series yourself.

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

The mechanism does; the numbers do not. Non-determinism, sampling error and window sensitivity are properties of how large language models generate text, not of any one language. The specific shares differ by language because the corpora differ, so use figures like these to calibrate what a noisy category looks like, not to estimate your own position.

Why does a category show a brand at 100%?

Almost always a thin brand set combined with few observations, not genuine dominance. Before treating any extreme figure as real, check how many brands are tracked in that category and how many samples went into the average. The conceptual issue underneath is which competitor set is being tracked in the first place — AI answers are assembled around the job the user is trying to do rather than around industry boundaries, a framing covered in this AIO fundamentals explainer.

Takeaway

GEO measurement is young enough that most teams are still arguing about which tool to buy. The larger source of error right now is upstream of the tool: reporting a noisy, low-sample metric over a window chosen after the fact. Fix the window, publish the sample count, state a minimum detectable change, and compare only like periods. It costs nothing, and it is the difference between a dashboard that tells you whether your work is paying off and one that tells you whichever story you went looking for.

Disclosure: AIOPulse is developed by Marketing Cloud Lab, which also operates this site. All figures cited from its public rankings were retrieved on 10 September 2026, except the 10 July to 9 August 2026 CRM window, which was retrieved on 10 August 2026. Competitor pricing for Profound, Peec AI, Otterly.ai and Mieruca GEO reflects publicly published rates as of September 2026.