We asked 4 AI engines the same questions 3 times each. Most brand recommendations didn't survive the repeat.

Finding 1: Only 31% of recommended brands appear every time

Across all question-and-engine combinations, the engines named 681 brand-recommendation slots. Just 214 of them — 31.4% — appeared in every repeat run of the same question on the same engine. The other ~69% of brands flickered: on the shortlist in one run, gone in the next.

If you've ever screenshotted an AI answer as proof your brand “is” or “isn't” recommended — that proof has roughly a one-in-three chance of describing the next answer.

Finding 2: Even the subject brand itself flickered in 1 of 4 cases

For the product we audited, we looked at every question-engine pair where it was mentioned at all: in 25% of those pairs, the product appeared in some runs of the identical question and vanished in others. Same engine. Same words. Different answer.

Finding 3: The engines disagree with each other — a lot

Same product, same 25 questions, wildly different visibility per engine:

A brand can be a fixture on Claude and a coin flip on ChatGPT simultaneously. Monitoring one engine — which is what a quick manual check does — tells you almost nothing about the others. We dug into that engine-by-engine gap across four businesses in which AI engine actually recommends your business: ChatGPT recommended the audited business least of the four, and its rivals most.

Finding 4: Consistency exists — at the top

It wasn't all noise. The product's two strongest competitors appeared in AI answers more consistently than the product itself — one was named in every single run of several questions. The engines do converge on a stable shortlist for each question; the brands on it benefit from a consensus across directories, comparison articles, and reviews that the engines see everywhere they look. The flicker zone is where everyone else lives — visible enough to appear, not established enough to stay.

That's the actionable part: AI visibility isn't a lottery. It's a frequency you can measure, and the gap between “flickering” and “fixture” is a list of sources you're missing.

Methodology

One subject product (web analytics; anonymized). 25 buyer-intent questions written from real search and forum phrasing. Each question submitted 3 independent times to ChatGPT, Perplexity, Gemini, and Claude via API on 2026-06-12 — 300 scheduled answers, 258 completed (engine rate limits cost some runs, OpenAI's the most). Every answer parsed for brand mentions by a judge model; “appears every time” = named in all completed samples of that question-engine pair (pairs with ≥2 completed samples counted).

What this means if you own a brand

Any verdict about your AI visibility based on one answer — yours or a vendor's — is statistically close to meaningless. Visibility is a mention rate, measured per engine, across repeats. That's exactly what an AskedAbout audit measures: 25 of your buyers' questions × 4 engines × 3 samples, with share-of-voice against every competitor the engines name.

Companion reading: which AI engine actually recommends your business breaks this run-to-run data down engine by engine. And if you're weighing a tool to track all of this for you, see the AI visibility tools compared — including the one-time alternative to a monthly subscription.