How much do AI recommendations move week to week?

What it means if you track AI answers

• Do not read one answer, or one week, as a change. A question that was named last week and not this week has, about one time in three, changed by chance alone.
• Read the rate, not the single answer. Each answer is one draw. Over many answers, a question settles at the share of times an engine names you, and that share is what moves when something real changes.
• Treat a move as real only if it is larger than the noise, or holds for several weeks. For 50 answers a week, that is a move of about 10 points or more, or a smaller one that lasts.
• Ask more to see more. More questions, or the same questions asked more often, shrink the noise with the square root of the number of answers.
• Judge each engine separately. Two engines asked the same question at the same moment often disagree, so one engine naming you says little about another.
• Keep failed answers out of the count. A call that failed is not a “not named”.

This is why every AEOTrack score carries a ± range and a “within noise” label (how the range is built), and why the free audit calls its one answer per engine a snapshot.

The data

What	Figure
AI answers read	8,758 for 52 businesses (55 tracked websites, 611 questions), 24 Jul to 19 Oct 2026 (14 weeks): 8,663 from the five engines and 95 from a two-week preview of Google AI Mode. 9,724 calls were stored; the 966 that failed are not answers.
Answers in the study set	7,192 usable answers with web search on, from 5 engines, to 514 questions that do not contain the business’s own name, 17 Aug to 19 Oct 2026 (10 weeks)
Answers that named the business	9.4% (678 of 7,192); by engine 8.4% to 10.8%
Businesses named at least once	27 of 52; the other 25 were never named in a web-search answer
Week-to-week flip, where that engine had named the business at least once	35% (99 of 283 pairs of adjacent weeks)
Two engines, same question, same check: when either named the business, both did	39%
Answers citing the business’s own site	6.5%; 57% of the answers that named the business cited it
Engine calls that failed	10.0% of the calls to the five engines over the period; 0.7% in October

How often each website was named, as a share of its own answers:

Share of a website’s answers that named it	Websites
0%	27
under 5%	11
5% to 15%	3
15% to 30%	6
30% to 50%	6
50% or more	2

How much the verdict moves

A series is one question on one engine. For each week we took the last check of that week, and compared adjacent weeks. A flip is “named” in one week and “not named” in the next, or the other way round.

Which series	Pairs of weeks	Flips (gains / losses)	Flip rate	95% range
All question-engine series	2,238	99 (49 / 50)	4.4%	2.3% to 6.9%
Questions some engine named at least once	523	99	18.9%	15.8% to 22.5%
This engine named the business at least once	283	99	35.0%	23.6% to 47.5%

The all-series rate is low because most series are never named at all: “not named” followed by “not named” is not a flip. The 35% is the figure that matters for a business that does get named. The ranges for the first and last rows resample whole websites, because the weeks of one account are not independent; leaving any one website out gives 32% to 41%.

By engine, where that engine named the business at least once:

Engine	Flip rate	Flips / pairs	95% range
ChatGPT	36.6%	26 / 71	26% to 48%
Claude	29.2%	14 / 48	18% to 43%
Gemini	35.6%	16 / 45	23% to 50%
Grok	33.9%	20 / 59	23% to 47%
Perplexity	38.3%	23 / 60	27% to 51%

The ranges overlap, so the data cannot rank the engines. Of the 249 series that were named at least once and have two or more weekly samples, 98 (39%) were named in some weeks and not in others.

Change or noise?

Four checks point the same way: most of the movement is noise.

1. Time does not matter. An answer asked again a few minutes later flips about as often as one asked again a week later.
2. The order looks random. 88 series have three or more weekly samples and at least one “named”. They show 77 flips, against 75 expected if the same verdicts came in a random order. There are no streaks.
3. Gains and losses balance: 49 against 50.
4. Model changes do not explain it. Pairs answered by the same model flip at 35.1% (where named), and pairs across a model change at 33.3% (only 21 pairs).

Under 1 hour31%

1 to 3 days22%

4 to 10 days32%

11 to 24 days45%

25 days or more27%

Time between the two answers	Flip rate	Flips / pairs
Under 1 hour (median 3.6 minutes)	31%	32 / 102
1 to 3 days	22%	28 / 126
4 to 10 days	32%	94 / 290
11 to 24 days	45%	13 / 29
25 days or more	27%	7 / 26

The “under 1 hour” pairs come from 11 websites, and one of them gives half of their flips; without it the rate is 23% (16 of 70). The longer gaps have few pairs, so their rates are rough.

This cannot prove that nothing real changed for any business. It shows that the movement we observed is what noise alone would produce.

Do the engines agree?

These figures compare two engines answering the same question in the same check, both with a usable web-search answer.

Questions	Pairs of answers	Agree	When either named the business, both did
All questions	13,218	91.6% (mostly “neither named it”)	39.3%
Questions some engine named at least once	3,282	66.9%	39.9%

By pair of engines, “both named when either did” runs from 31% (ChatGPT and Gemini) to 44% (Claude and Grok). On the questions some engine named at least once, the same engine agreed with its own answer from the week before 81% of the time, while two different engines in the same check agreed 67% of the time.

1 of 593

2 of 555

3 of 532

4 of 518

All 538

Of 1,174 checks where all five engines answered, 236 had at least one engine naming the business. In 39% of those only one engine of five named it, and in 16% all five did.

How many answers before a change is real

Each answer is one draw: the engine names the business or it does not. For a question an engine names you on some of the time, that draw is a weighted coin, and the weekly results above behave exactly like one. So the usual arithmetic for coin tosses applies: one draw tells you almost nothing, and the noise shrinks with the square root of the number of draws.

One question on one engine, one answer a week

If a question was named in half of its weekly answers, the range in which its true naming rate lies (95%) is:

Weekly answers	Range for the true naming rate
4	15% to 85%
13 (a quarter)	26% to 74%
52 (a year)	37% to 63%

Telling a question named half the time from one named a fifth of the time takes about 18 weekly answers on each side to clear the noise, and about 36 to be fairly sure (80% power). One question’s “named last week, not this week” should never be reported as a change.

A whole account, asking the same questions again

With no real change, the week-to-week move in an account’s share of answers that name it has a 95% noise band of about ±1.96 × √(f ÷ n) points of share, where n is the number of answers a week and f is the share of question-engine series that flip from one week to the next. From the data, f is 4.4% across all series, about 12% for the median account that does get named (7 websites named in 10% or more of their answers, range 7% to 20%), and 35% where the engine named the business at least once.

Answers a week	f = 4.4%	f = 12%	f = 20%	f = 35%
10	±13.0	±21.7	±27.7	±36.7
50 (10 questions × 5 engines)	±5.8	±9.7	±12.4	±16.4
150 (the same 50, each asked on 3 days a week)	±3.4	±5.6	±7.2	±9.5
250	±2.6	±4.3	±5.5	±7.3

Averaging k weeks before comparing narrows the band by √k, and asking each question on 3 different days a week narrows it by √3, about 42%. (Asking the same question again in the same second measures nothing: why we ask once per check.) For an account with 50 answers a week, the weeks needed on each side are:

Change to detect	f = 12%: clears the noise	f = 12%: fairly sure	f = 20%: clears the noise	f = 20%: fairly sure
5 points	4 weeks	8 weeks	7 weeks	13 weeks
10 points	1 week	2 weeks	2 weeks	4 weeks
20 points	1 week	1 week	1 week	1 week

Checked against the data. For 11 account-weeks on 8 websites that do get named (a median of 50 matched answers), the actual week-to-week moves had a median of 4.0 points, a 90th percentile of 7.5 and a maximum of 15.4, against a predicted noise band with a median of ±8.0. Measured in units of that band, the moves had a spread of 1.02, about what pure noise gives, and 1 of the 11 fell outside it. Across all 42 account-weeks with 20 or more matched answers (25 websites), the median move was 0 and 2.4% moved 10 points or more, because most accounts are never named and their share sits at 0.

When the two periods use different answers (no repeated questions), the textbook count of answers needed in each period is:

Change in naming rate	Answers per period to clear the noise	Fairly sure (80% power)
5% to 10%	212	432
10% to 20%	97	197
20% to 30%	143	291
30% to 40%	173	354
50% to 60%	189	385

Own-site citations and failed calls

Own site cited. 6.5% of answers in the study set (468 of 7,192) listed the business’s own site among their sources, against a median of 9 sources per answer. Of the answers that named the business, 57% cited its site: Perplexity 74%, ChatGPT 69%, Claude 47%, Gemini 46% and Grok 44%. Gemini lists sources on only about half of its answers and Claude on 78%, so their rates are floors. 82 answers cited the site without naming the business, 49 of them from Perplexity. Being read is not the same as being named.

Failed calls. 966 of the 9,629 calls to the five engines failed (10.0%; the 95 calls to the Google AI Mode preview all answered), and 27% of checks were missing at least one engine. Failures clustered: one day in late September held 47% of them; without it the rate was 5.7%. By month: July 20.0%, August 16.8%, September 12.4%, October 0.7%. 932 of the 966 were requests the provider rejected (HTTP 400 to 404), which points to key, credit or model-name problems on our side rather than to the engines; only 34 were the provider being unavailable. A failed call is our own history, not a measure of an engine’s reliability, and it is never counted as “not named”. One automatic retry was added on 12 October, so rates from that date undercount first attempts.

Method

• Who is in it. Real accounts only: test accounts (including our own self-tracking site) are left out, and so, on privacy grounds, are two further websites. Businesses are counted by site address: 55 tracked websites make 52 businesses, because three were signed up twice.
• Which answers count. The five engines (ChatGPT, Claude, Gemini, Grok and Perplexity), asked through their APIs with live web search on, in a fresh session; not failed; not cut off at the length limit. Left out of the 8,758: 928 answers without web search (24 July to 13 August, before the engines were switched to their search APIs), 490 cut-off answers and the 95 Google AI Mode preview answers, which leaves 7,245. The 1,971 answers to a different kind of check (topics) are not among the 8,758. The 4 questions that contain the business’s own name (53 answers) are left out of every “named” figure.
• What “named” means. The verdict of AEOTrack’s matcher (what counts as a mention), applied to every stored answer with the same rule, so no flip comes from a rule change. Two corrections from a manual audit on 21 October are applied: 48 answers the matcher missed are set to named, and 59 “named” verdicts for an account matched to a different company of the same name are cleared.
• Weekly series. One sample per question and engine per ISO week (Monday to Sunday, UTC): the last check of the week. Using the first check of each week instead gives the same figures (4.4% and 35.6%).
• Uncertainty. Ranges are Wilson 95% intervals. The headline flip rates also resample whole websites 2,000 times, because answers from one account are not independent. The randomness test shuffles each series’ weekly verdicts 2,000 times; 77% of the shuffles had as few flips as observed, or fewer.

The headlines barely move under other choices:

Variant	Flip (all)	Flip (named at least once)	Named share	Businesses named	Both named when either did
As published	4.4%	35.0%	9.4%	27 of 52	39.3%
Stored verdicts, without the two corrections	4.7%	34.8%	9.7%	28 of 52	37.8%
Including answers without web search	4.4%	33.9%	8.9%	29 of 52	37.7%

Caveats

• Small and short. 52 businesses, most of them tracked for a few weeks: the median series has 2 weekly samples (90th percentile 3, maximum 7). The 35% rests on 283 pairs from 18 websites; the largest website gives 18% of those pairs and 12% of all study answers.
• Per-engine figures use 45 to 71 pairs each. Do not rank engines on them.
• Not a random sample. These are businesses, mostly small, that signed up to an AI visibility tool. Our product suggested most of the questions, and some of them poorly. The figures describe these questions, not “AI in general”.
• API answers, not the consumer apps. The answers came from the engines’ search APIs, almost all without a country set. The apps people use remember, personalise and localise, so they can answer differently (why your own chat may answer differently).
• The matcher is automatic. Its errors were audited but are not zero: the 48 corrected misses are a floor, and before the corrections a manual sample of 81 answers agreed 92.6% of the time.
• “Named at least once” is chosen by its outcome. That raises flip rates in short series, which is why the all-series and per-question rates are given beside it.
• Within a single check we could not measure variation: the 36 checks that asked more than once never named anyone.

Privacy

Every figure here is an aggregate across accounts. The page names no customer, website, question or answer, and none can be worked out from it. Accounts flagged as test accounts are left out, and so are two further websites, on privacy grounds. The data was read from a copy of our database taken on 21 October 2026, read-only.

Questions about the data or the method go to support@aeotrack.io, which Swapnil Patle, Product Head, answers. To cite this page: AEOTrack, “How much do AI recommendations move week to week?”, 21 October 2026, https://www.aeotrack.io/research/ai-answers-week-to-week.

Related reading: How the AI visibility score is calculated · One answer per engine: what the ± means · Why your score changes