AI visibility is measured on one persona

Persona conditioning is the effect a buyer's stated context (their role, company size, budget, tooling preferences) has on which brands an AI assistant recommends for an otherwise identical prompt. It is large, it is uneven across brands, and almost no visibility tracking accounts for it.

#The audit

A cross-provider audit published in May 2026 ran the same commercial prompts through OpenAI and Anthropic models with ten different buyer personas prefixed to the user message. The design was 2,000 trials: ten personas by eight prompts by three model configurations by ten repetitions. The measure was Jaccard similarity between recommendation sets, meaning the share of recommended brands two answers have in common.

Repeating the same persona produced within-persona similarity of 0.42 to 0.51, while cross-persona similarity fell to 0.22 to 0.35. The persona effect itself came out at a delta of -0.12 to -0.20 across the three cells, and it held with clustered confidence intervals: GPT-5.4-mini at low reasoning gave -0.12, the same model at high reasoning gave -0.16, and Claude-sonnet-4.6 at low gave -0.20.

Two things in those numbers matter more than the headline.

The first is the baseline. Even holding the persona fixed and asking the same question ten times, the recommendation sets only overlapped about half the time. Persona conditioning is a shift stacked on top of an already unstable answer, not a clean signal on a stable one. We have written before about why AI visibility is a distribution, not a number, and this is the same finding arriving from a different direction.

The second is that the effect is not distributed evenly across brands.

#Prominence decides how much you swing

The audit stratified brands by prominence and reported swap rates by stratum. Category leaders were the most persona-resistant, changing 20% to 29% of their recommendation slots across personas. Regional players sat close behind at 27% to 33%. Mid-market brands were the outlier, with 39% to 75% of the recommendation set replaced as the persona changed.

That is the practical takeaway for most companies reading this. If you are the default answer in your category, your tracked visibility number is roughly the number your buyers see, because your position survives most contexts. If you are in the middle of your market, up to three quarters of the answer can turn over depending on who is asking, and a single tracked score is describing one buyer out of many.

It also reframes what a visibility drop means. A mid-market brand that loses ground in a tracker may not have lost a ranking. It may have been measured against a persona it was never going to win, in a week when the tracker's prompt set drifted.

#Some of the recommendation never touched retrieval

The same audit logged whether each recommended brand could be traced back to a document the model actually retrieved. It could not, often. Anthropic configurations produced 43% to 52% of recommendations with no observed retrieval-layer evidence, against 8% to 29% for the OpenAI configurations.

Recommendations with no retrieval trail come from the model's parametric knowledge, not from your page. No amount of on-page work reaches them in the moment. They move on a much slower clock, driven by what the wider web said about you before training, which is a reason to care about third-party coverage rather than only about your own documentation.

#The field knows this is a gap

This is not a fringe concern. The critical survey of generative engine optimization, covering 45 studies published between November 16, 2023 and July 14, 2026, names personalization as an open research gap rather than a settled area. Its own summary of the corpus is that most studies use English, anonymous accounts, and a small number of locations.

Anonymous accounts are the default because they are reproducible. Tracking tools open a clean session precisely so that yesterday's run and today's run are comparable. That choice buys you a stable time series and costs you external validity: you get a reliable measurement of the one buyer nobody actually is.

#The memory layer is already deployed

Anonymous measurement was defensible when assistants were mostly stateless. They are not anymore. OpenAI shipped a rewrite of the ChatGPT memory architecture on June 4, 2026, with roughly 5x lower compute for free-tier memory synthesis and 2x memory capacity for Plus and Pro. The important change is not capacity. It is that memories are now synthesized automatically from chat history rather than saved on explicit request, so a returning buyer carries context they never deliberately set.

Product surfaces are moving the same way. A preference-memory reranker for shopping agents reported gains of up to 10.61 absolute points in 1-in-5 accuracy over no-memory and raw-history baselines by distilling purchase history into compact, query-independent signals. Personalization is being engineered into the recommendation path because it works, which means the persona effect measured above is a floor, not a ceiling.

#What to change in your measurement

The audit is small: eight prompts, two providers, one product category framing. Treat the direction as real and the exact percentages as provisional.

Three changes follow from it anyway.

Run personas as a deliberate axis, not as noise. If your buyers split into three or four recognizable types, that is three or four prompt sets, tracked separately. A blended average across them hides exactly the variance that matters, and for a mid-market brand the average is dominated by the segments where you already lose.

Report a range instead of a score. Given a within-persona overlap near 0.5, any single-run number carries more uncertainty than a dashboard implies. The same logic applies to how many repeats are worth paying for: we covered the diminishing returns in the fifth rerun buys almost nothing, and adding a persona is a better use of the same budget than adding a repeat.

Separate the two levers. For recommendations backed by retrieval, page-level work is in scope. For the large parametric share, the lever is third-party presence, and the feedback loop is measured in quarters. Mixing them into one visibility score makes both look unresponsive.

The prompt "best CRM software" was never one query. It is a family of queries that differ by who is asking, and your number so far has only answered one member of it.