How much do AI answers change from one week to the next?

Not exactly, in our small sample. We asked the same 9 questions a week apart. Of the tracked tools named or linked in either answer of a pair, just over half appeared in both (35 of 60). Of the sites the answers cited or linked, about a third did (96 of 277, ChatGPT and Perplexity only). It’s two dates and 9 questions, so read this as what we saw, not as a rule.

Those are pooled counts. A pair is one question, on one assistant, in one wording, answered on two dates. For each pair we counted the tools in either answer and the tools in both, then added the pairs up: 35 of 60 across 26 pairs. A tool that turns up in several pairs is counted in each of them, so that total is not a count of different tools. The same goes for sites. Neither figure is a total of unique tools or sites.

If you check your own business by hand, this matters: one answer is one sample. This post shows what we saw in our own tests, how we counted it and what to do about it.

Will ChatGPT give me the same answer next week?

Not reliably, in our sample. Across all three assistants, the exact same set of tracked tools came back in only 3 of the 15 pairs where at least one was named or linked. In the other 12 pairs, the sets differed. That doesn’t make one answer right and another wrong. It means a single answer is one draw and the next draw can differ.

For ChatGPT in particular, 11 of 17 tracked tools named or linked appeared in both answers a week apart and 9 of 33 cited or linked sites did. Those counts are small, as the section on assistants below explains.

What did we compare and how?

We compared two Answersky tests and one repeat of the second. The comparison uses only the wording both tests share.

• Sep 30, 2026 test: 9 questions × 3 wordings × 3 assistants (ChatGPT, Gemini, Perplexity), each asked once. We use the main wording only: 27 answers, of which 26 were usable because 1 Gemini answer came back empty.
• Oct 7, 2026 test: the same 9 questions in the main wording × 3 assistants × 2 repeats. That is 54 answers and all 54 were usable.

That gives three comparisons. Sep 30 against the first Oct 7 repeat is a week apart, with 26 usable pairs. Sep 30 against the second Oct 7 repeat is the same again, as a check. The two Oct 7 repeats against each other are minutes apart, with 27 pairs. We asked the assistants through their APIs with web search, so the apps themselves can answer differently.

What “named or linked” means. For tools, we track 8 other tools that track AI visibility. A tool counts if its name is written in the answer, or if its website is among the addresses the answer cited or linked, or is in an address written out in the answer text. So we say “named or linked”. Almost all of it is names written in the answer: counting only those gives 33 of 57 a week apart and 41 of 62 the same day. 7 of the 8 appeared in the Sep 30 test and the same 7 in the Oct 7 test. So this measures how steady that list is, not how steady every business named in an answer is.

Answersky is counted separately. For us, “named” means our name in the answer, or answersky.com among the cited or linked addresses. Neither appeared in any answer in either test.

What “cited sites” means. The website part of each address an answer cited or linked, plus any address written out in its text. Gemini returns its sources as redirect addresses that we don’t resolve yet, so we leave its cited sites out of every comparison. Tracked tools named or linked use all three assistants.

How many tracked tools were still named or linked a week later?

In our sample, just over half of the tracked tools named or linked in either answer came back after a week and about two thirds came back within minutes. The figures are pooled across pairs, as above.

• A week apart: 35 of 60 across 26 pairs. The second Oct 7 repeat gave 35 of 57.
• The same day, minutes apart: 45 of 65 across 27 pairs.

The gap between a week and a few minutes is small for so few questions, so we don’t read much into it.

ChatGPT named or linked a tracked tool in only 3 of its 9 answers on Sep 30 and in 3 of 9 on each Oct 7 repeat, so many of its pairs had no tools to compare.

How many cited sites repeated?

Fewer than the tracked tools did. This covers ChatGPT and Perplexity only, for the reason above. Pooled across 18 ChatGPT and Perplexity pairs:

• A week apart: 96 of 277 cited or linked sites appeared in both answers, about a third.
• The same day: 181 of 224 cited or linked sites appeared in both answers.

Those two totals combine assistants that behave very differently, which is the next question.

Which assistant changed most?

It depends on what you count. In this sample no assistant was steady on everything.

• Perplexity’s tracked tools named or linked matched in 13 of 15 within the day and in 9 of 18 across the week.
• Perplexity’s cited or linked sites matched in 176 of 185 within the day and in 87 of 244 across the week. We don’t know why.
• ChatGPT’s cited or linked sites matched in 5 of 39 even minutes apart and in 9 of 33 a week apart. ChatGPT listed at least one site in 6 of its 9 answers on Sep 30 and in 7 of 9 on each Oct 7 repeat, so these are small counts.
• Gemini’s tracked tools named or linked matched in 15 of 25 across the week and 21 of 34 within the day.

We wouldn’t rank the assistants from this. Each one moved and each moved differently depending on what we counted.

What changed in our method between the tests?

Three things. We handled each one and the comparison still has limits.

• The tool matcher was corrected before the Oct 7 test. It had under-counted some tracked tools, so we recounted both tests with the same, corrected matcher for this post. Whether Answersky was named did not change.
• The tests asked different sets. Sep 30 asked 3 wordings once each. Oct 7 asked one wording twice. We compare only the wording both tests share, which leaves 9 questions.
• Gemini’s cited sites are not counted, as above.

Two tests a week apart is still a small sample. We will post later tests in our public log.

What does this mean for checking by hand?

Treat one answer as one sample, then repeat it and compare. A name that shows up once and not the next time is not a trend and neither is one that disappears.

• Save each answer with its date. A chat that scrolls away can’t be compared later.
• Ask each question more than once and in a few wordings, which is how our own tests are built. We explain why on how it works.
• Compare lists, not single lines. Note which tools and sites come back most often.
• Repeat the check on another day before you act on a change.

Our post on checking if ChatGPT mentions your business gives a sheet you can copy and the one on checking your AI visibility yourself with Claude has prompts to start with.

Does Answersky show up yet?

No. On Sep 30 we were named in none of 78 usable answers (all 3 wordings) and on Oct 7 in none of 54, in the answer text or in the addresses it cited. That is the same result twice. Around us, some of the tracked tools and sites moved, as above.

We make no promise that the next test will look different. We will report what it shows either way.

Free Quick Scan: 6 of your buyers’ questions on ChatGPT and Perplexity, 12 answers and who AI named instead of you. Results on the scan page when ready, plus an emailed copy; queued or review requests arrive by email within 1 business day.