Google’s AI Knows This Working Actress in Detail. She Has No Wikipedia Page.

Casting is a fully answered question

Every one of the twenty queries produced an AI Overview. Not most — all of them. “Best rising young actresses,” “up and coming actors to watch,” “best actors to play a villain,” “best bilingual latina actresses”: Google writes a substantive answer, names people, and lists its sources every time.

That is worth setting against the companion study we published this week. When we ran 72 local automotive service searches through the same surface, an AI answer appeared 6.9% of the time. Same engine, same week, same country. Ask it who should fix your car and it mostly declines; ask it who to cast and it always has an opinion.

Who Google reads about actors

Share of the 20 answers citing each source:

IMDb is the spine, as expected. The surprise is at the bottom: Backstage — the platform the industry actually uses to cast — was cited in one answer out of twenty, behind Facebook, YouTube, Reddit, Quora and Teen Vogue. When an AI is asked who to cast, it is reading a filmography database, social platforms and consumer magazines. It is largely not reading the casting industry’s own publications.

Across the twenty answers there were 78 distinct source domains and 150 citations, with 61 of those domains cited exactly once.

The worked example: no Wikipedia, described anyway

The assumption going in — ours, and the industry’s — was that an AI can only describe a performer who is documented in the canonical reference sources. No Wikipedia article, no IMDb depth, no visibility. We tested it on a real working actress rather than a hypothetical.

Gredivel Vasquez is a bilingual Venezuelan-American actress based in New York, a long-time company member at Repertorio Español, and a SEMPITE client — disclosed here because a study that names a real person should state plainly where the relationship lies. She has no English Wikipedia article. We checked directly against the Wikipedia API: the page does not exist, and a site-wide search for her name returns zero results.

Asked who she is, Google’s AI Overview answered:

A second query returned a dated filmography — Me Fui con Ella (2026), Capturada (2025), Your Monster (2024) as Diana Franco, A Moment for Glory (2023) — and a third listed her stage roles by production and character, along with a 2025 Premios Talía award for Best Actress in Drama.

That is not the output of a system that has never heard of someone. It is a competent professional summary, and it is accurate.

Her work is at gredivelvasquez.com — the same page Google’s AI read to answer the question.

Why this contradicts what a model will tell you

Run the same question past a conversational model working from training data and you get the opposite result: no reliable knowledge of her, followed by a confident list of famous names. That is the version of this test most people run, and it produces a bleak and misleading conclusion — that working actors are invisible to AI and only the famous exist.

The distinction is between a model recalling what it memorised and a surface reading the live web. Google’s AI Overview does the latter, which is why it can describe a performer with no encyclopedia entry, and why it named genuinely emerging actors in our discovery queries — Milly Alcock, Sophie Wilde, Yerin Ha, Tyriq Withers, True Whitaker, Tanzyn Crawford — rather than recycling a list of stars.

If you take one thing from this study, take that: on live-sourced surfaces, the barrier to being described accurately is not fame. It is legibility.

No, it is not the same famous names on repeat

The assumption this study set out to test was that an AI asked who to cast will recycle a short list of stars. Across the twenty answers, Google named 101 different performers in 119 mentions — an average of six per answer.

87 of those 101 were named exactly once. Only four people appeared in three or more answers:

The ten most-repeated names account for just 20.2% of all mentions; the top three, 7.6%. For comparison, in our supplements study a single brand was named in 61.5% of answers. There is no equivalent here. On this surface, casting answers are a long tail, not a star system.

Where it still breaks: the bilingual queries

The picture is markedly worse on our second bucket of eight queries about Latino and bilingual performers.

Asked for the best Mexican-American actors, Google’s answer profiled Anthony Quinn, who died in 2001, and Lupe Ontiveros, who died in 2012, alongside living performers — an answer a casting director could not act on. Asked for best bilingual latina actresses, it returned Salma Hayek, Penélope Cruz, Sofía Vergara and Eiza González: accurate, famous, and not a discovery list. And three queries — rising afro latina actresses, best latino actors in hollywood, and best spanish speaking actors in american tv — produced answers with no profiled performers at all.

Compare that with the general-discovery queries, where the same engine surfaced six emerging names in a single answer. The capability plainly exists. It is being applied unevenly, and the unevenness falls on exactly the performers with the least existing coverage to draw on.

Methodology

Data collected 6 August 2026, United States, English, desktop. Twenty search-shaped casting and discovery queries — twelve general, eight bilingual/Latino — submitted to Google through the DataForSEO live SERP endpoint with AI Overview loading enabled. Every query that errored was retried until it returned; there are no unrecovered failures. For each answer we recorded the AI Overview body text and every domain in its reference block. The worked example was verified separately on the same surface on 6 August 2026, and the Wikipedia result was confirmed directly through the MediaWiki API rather than inferred from a search page.

How performers were counted. Google names people two ways: in a structured “Name: description” profile block, and in a lead sentence roster (“…include Stephen Root, Margo Martindale and Bill Camp”). Extraction is rule-based and handles both, so it can be reproduced exactly: profile names are taken from the start of a sentence immediately preceding a colon, and roster names are parsed from the first sentence after a trigger word, with parenthetical titles stripped. Film and series titles never occupy either slot. The extractor is validated against four answers whose casts we labeled by hand from the full text, and the pipeline refuses to emit results if any disagree — a check that caught one of our own hand-labels built from a truncated printout. It resolves names in all 20 answers. An earlier version of this study withheld these figures because it recognized only profile blocks and therefore saw nothing in 11 of the 20 answers.

Limits. One snapshot; AI Overviews are non-deterministic and vary by user, location and time. This measures Google’s AI Overview only — not Google AI Mode, ChatGPT, Perplexity or Gemini, which are different surfaces with different sourcing, and a conversational model will answer these questions very differently, as noted above. Twenty queries is a probe, not a census. Career details in the worked example are as Google’s AI stated them; we quote the engine’s characterization rather than independently auditing every credit. And one bare-name query (“Gredivel Vasquez” alone) returned no AI Overview at all, while the three question-shaped variants did — phrasing changes whether this surface engages.

Query list, per-answer citation records and the worked-example capture are available on request under CC-BY.