We asked AI the same questions about 20 businesses two weeks later. All 20 answers had changed.

Finding 1: Not one business had a fully stable result

We define a result as two things: the AI Visibility Score (how often, and how prominently, the engines name you when your buyers ask) and the competitor list (the rival businesses the engines named instead). Across the 20 re-runs, 20 of 20 changed on at least one of those two — the score, the rival set, or both. Zero were unchanged.

This is the number we did not expect. We assumed a well-established business asking an identical question would mostly get an identical answer a fortnight later. That happened exactly zero times.

Finding 2: Half the scores moved — some a long way

The two largest moves were both upward and both large: a Dallas med spa went from 50 to 83 in 14 days, and a Fort Lauderdale personal-injury firm went from 17 to 42 in 23 days. Neither hired us; neither, as far as we know, did anything at all. The engines simply changed their minds. The two decliners were milder (47→41 and 28→26), but a business watching only its own score would have read those as noise.

The direction of travel is worth naming honestly: 8 of the 10 moves were upward. Some of that is the general expansion of what these engines retrieve about local businesses. It does not make the number stable — it makes it a moving target that happens, on this sample and in this window, to be drifting up.

Finding 3: A flat score does not mean a stable position

Ten businesses scored exactly the same both times. It would be reasonable to call those ten stable and move on. They weren't. All ten of them had a different competitor list the second time. Same score, different rivals standing in front of them — a different set of businesses the engines now consider the answer to their customers' question.

If you check your AI visibility once, see a mediocre score, check again and see the same mediocre score, the intuitive read is nothing is happening. On this data, that read was wrong in 10 cases out of 10.

Finding 4: 57% of the competitors were new names

Each check records the top 5 rivals the engines named. Across the 20 businesses that is 100 competitor slots per round. Comparing round one to round two:

• Only 43 of the 100 names survived. Fewer than half the rivals the engines named as the answer to a buyer's question were still named two weeks later.
• 57 of the 100 names in round two had not appeared in round one. Not lower-ranked — absent.
• Average overlap between the two rival lists: 0.29 (Jaccard). Two lists of the same length, describing the same market, on the same day of the week, agreeing on roughly a quarter of their contents.

The competitive picture is substantially less stable than the score. That matters, because the competitor list is the part of an audit people act on — it is where you look to find who has the review volume, the directory listings, and the roundup mentions you don't. Half of that list has a shelf life measured in weeks.

Finding 5: Being named as often is not being named the same

Of the 10 businesses whose score moved, 3 were mentioned in exactly the same number of AI answers as before. Their mention rate was identical; what changed was where they landed inside the answer — named first with a reason, versus listed fourth in a trailing "others include…" clause. The engines had not stopped naming them. They had stopped recommending them.

Mention-counting alone would have shown a flat line through all three of those cases. This is the same lesson our run-to-run study found inside a single day — a screenshot proves nothing — extended across weeks: a mention count is not a position.

Finding 6: The lower-visibility vertical moved more

The law firms are the smaller sample and we won't over-read it, but the pattern is consistent with what we see across the full local-business benchmark: the more fragmented and contested a category is — 190 distinct rival firms named across just 10 personal-injury audits — the more freely the engines reshuffle it. Nobody in that category has become the consensus answer, so every re-run is a fresh election.

Methodology

Drawn from our production audit corpus: 134 completed AI visibility audits on 93 distinct US businesses, 790 engine answers, 463 distinct rival businesses named, run between 2026-06-12 and 2026-07-09. From those we took every business audited twice, at least 7 days apart, where the question inputs were byte-identical — same business name, same vertical string, same location string, and therefore the same three generated prompts. That is n=20, with gaps of 14 to 23 days (median 14). We discarded 19 other re-audits whose vertical or location text had been edited between runs, because their prompts were no longer literally the same question.

Each audit sends three buyer-intent prompts ("What are the best {vertical} in {city}?", "I'm looking for {vertical} in {city}, who do you recommend?", "Is {business} a good choice?") to ChatGPT (OpenAI) and Perplexity — six answers per audit, 240 answers across the study. A judge model parses every answer for whether the subject was mentioned, how prominently, and which rival businesses were named. The score is the mean of a per-answer prominence weighting, 0–100.

What we cannot claim: we did not contact these 20 businesses and we do not know whether any of them changed their marketing, earned press, or gained reviews during the window. Some of the movement is certainly real-world change. What we can say precisely is that nothing about the question changed, and that a business owner re-checking on our schedule would have received a different answer in 20 cases out of 20. Individual figures are rounded. The subject businesses are audited from public information and are described here by metro and vertical only — none is a customer, and we do not publish individual businesses' scores against their names.

What this means if you own a business

The honest conclusion is narrower than "check constantly." It is this: an AI visibility number has a half-life, and on this evidence that half-life is shorter than a month. A single check is genuinely useful — it is the only way to find out whether the engines name you at all, and which rivals they name instead. It is a photograph, not a fixed address.

So use it as a baseline, and re-measure after you change something — new directory listings, review volume, a comparison page, a roundup mention. A change you can't measure against a baseline isn't a fix, it's a hope. And if your score is flat, look past it at the rival list: on this data that list moved every single time, and it is where the actual instructions live.

Related reading: do AI answers change every time you ask? (yes — 69% of recommended brands didn't survive a same-day repeat), which AI engine actually recommends your business, and why AI recommends your competitors instead of you.