A visibility baseline from 15 questions across four AI engines

We sent 15 saved questions about finding individual marketing experts to ChatGPT, Gemini, Claude and Perplexity. The 15 October 2026 collection returned 60 answers. None of the returned text contained the participant's exact saved name.

After collection, we found that the first extractor missed names in several saved answer formats. We fixed those formats and reanalysed the same 60 saved answers. This was not a rescan. We made no new requests to the four engines. The original analysis remains in the evidence archive.

What we tested

The saved topics were Performance Marketing, Google Ads and Meta Ads, with Europe as the professional geography. Each topic had five questions covering hiring, solving a business problem, learning, finding a speaker and comparing practitioners.

These profile-derived questions were already saved in the platform. They are not measured search demand, and none included the participant's name. One hiring question asked, “Which individual independent consultants specialize in Performance Marketing in Europe?”

Collection ran on 15 October 2026, from 11:41:45 to 11:50:39 UTC. Each engine received each question once. ChatGPT and Gemini were collected as web answers with US English requested. Claude and Perplexity used API answer sampling. Europe in the question wording does not prove a European search location.

The corrected analysis

The first analysis used version 1. The corrected analysis uses version 2, recorded at 2026-10-15 12:58:47 UTC. It handles the numbered and bold candidate formats found in the saved answers. It keeps a source with a person only when an explicit URL in that local entry matches a saved source. Numeric citation markers stay at answer level because source identifiers were discarded and duplicate URLs were merged during earlier normalization.

“Automatically eligible” means a returned, nonempty discovery answer with a usable or limited analysis and a recorded identity assessment. Non-answers and excluded answers remain unmeasured rather than counted as missed mentions. No Gist Score is reported.

Across all answers, the first analysis recorded 4 usable, 29 limited and 27 unusable answers. The corrected analysis recorded 3 usable, 48 limited and 9 unusable answers. These labels describe this extractor's reading of saved responses. They are not a ranking of engines or people.

The raw response text, citations, timestamps, answer identifiers and collection method remain unchanged. The evidence hash is stored privately with the research record. Reanalysis changes only the analysis fields and preserves the original version for comparison.

What the answers do and do not show

An exact-name search found no appearance of the participant's saved name in the 60 returned answers. That observation applies to this profile, these questions and this collection window. It does not show that the participant is invisible to AI, because it does not test aliases, every possible question or another user's personalised session.

Some answers gave directories or search advice, asked for more context, or requested app access instead of candidates. Those remain unmeasured rather than evidence of a missed mention. The corrected parser should not turn a non-answer, an organisation or an abbreviated surname into a person.

The archive retains returned citation links at the answer level. We have not independently verified the credentials, experience, clients or results stated about candidates. A link in an answer does not prove every claim about a named person. Open cited sources before relying on a claim. OpenAI's search guidance likewise advises readers to check whether citations support an answer.

What this baseline supports

This is a dated archive of four collection channels answering a fixed question set. It supports a comparison between two analysis versions of unchanged saved evidence. It does not identify ranking factors, explain an absence or show that publishing content would change future answers.

There was no content intervention, control group or repeated sample. We cannot compare this collection directly with the older 24 September scan because it used a different method. A later rescan should preserve the question wording, engine scope and collection conditions. Gistful's methodology guide explains why changed scope can prevent a direct comparison.