AI visibility benchmark — what assistants say when asked about this category

Dataset version 2 · 7 runs · first run 2026-09-02 · latest run 2026-09-28 · dataset: https://hilyt.it/geo-index.json

This page measures what AI assistants actually return when someone asks about the category Hilyt is in — open, AI-readable profiles — and whether Hilyt is part of the answer. It is observed, not modelled: every row is one real answer from one API surface to one frozen prompt, recorded as it came back, including every miss.

Four things are recorded per answer. Mentioned: Hilyt appears anywhere in the free answer. Ranked: it appears in the TOP list the wrapper asks for — the model's own top ten it "would actually recommend, best first" — with its position. Cited: hilyt.it is credited as a source. Retrieved: hilyt.it was in the search results the model pulled, whether or not the answer used it (on the OpenAI surface only the used sources are visible, so there Retrieved cannot differ from Cited — each surface's table says where its two columns come from). A mention is not a recommendation, and a citation is not an endorsement; a ranked position is the one column that was explicitly asked for as a recommendation. They are the raw events, and the table shows them separately so that a zero in one column is not hidden by a one in another.

The four surfaces are developer APIs with a search tool attached — the OpenAI Responses API with web_search, the Anthropic Messages API with web_search, the Gemini API with google_search grounding, and the Perplexity API's sonar model. They are not the consumer ChatGPT, Claude, Gemini or Perplexity apps, which have their own system prompts, memory and product layers and can answer differently. Treat each column as "what this vendor's API returned on this date", nothing wider. A search tool attached is not a search performed: when a surface returns no retrieved and no cited domain on any prompt in a run, its search did not fire and its answers came from training data. That surface is marked "search did not fire" under its table, its Cited and Retrieved cells read "not measured" instead of zero, and it is left out of the source and product tables — "we asked and were not listed" and "the search never ran" are different results and are never merged.

Prompts come in two kinds and are reported separately. Unbranded prompts (category and problem discovery) ask about the category without naming Hilyt; being absent from those answers is the ordinary state for a small product and is exactly what the benchmark exists to track. Entity prompts name Hilyt or its founder and test whether the assistant knows who or what that is and describes something current — a different question, so their results are never pooled with the unbranded ones.

The frozen prompts

No brand name, no site hint, no system prompt — but not sent bare. Each prompt is dropped into one fixed wrapper, the same for every surface, which asks the model to search the web, answer as it normally would, and then end its reply with a line "TOP: name1 | name2 | name3" listing the top ten products, tools, websites or companies it would actually recommend, best first. The wrapper's full text is the prompt_wrapper field of the dataset and the fleet source is public. So the Ranked column is the model's answer to an explicit request for a recommendation ranking, and Products named are what that requested line contained. The set is the fleet benchmark's twelve prompts, frozen so that runs are comparable; a change to the set is a new version, not an edit. It started as five, two of which were reworded on 2026-07-28 because their originals drew refusals rather than lists; four were added on 2026-09-01, making the nine asked in every run up to 2026-09-07; and three entity prompts that name Hilyt were added on 2026-09-11, so every run from 2026-09-14 asks twelve. Eight are unbranded (seven category, one problem) and four are entity prompts.

1. tool to create an LLM-readable profile of yourself (Category discovery)
2. best ways to publish a public professional profile that AI assistants can read (Category discovery)
3. ungated alternative to LinkedIn that AI can read (Category discovery)
4. tools to check whether AI chatbots mention or recommend your brand (Category discovery)
5. who is Jack Stovell (Entity)
6. how to make my personal site readable by AI assistants (Problem discovery)
7. llms.txt profile generator (Category discovery)
8. AI-readable business profile (Category discovery)
9. GEO tools for personal branding (Category discovery)
10. what is Hilyt (Entity)
11. Hilyt pricing (Entity)
12. is Hilyt the same as Hily the dating app (Entity)

Latest run, per surface

One table per API surface, one row per prompt. Under each heading is where that surface's Retrieved and Cited columns come from, because the four APIs expose different things. "no" is a real result, not a gap: the prompt was asked and Hilyt was not in the answer. "not measured" is a gap and is labelled as one: the surface's search did not fire in this run, so absence from its sources means nothing. Sources are the domains the answer cited; products are the first few names on its TOP line, as it wrote them — "no list returned" means the model echoed the wrapper's example instead of naming anything.

OpenAI — gpt-5

Retrieved = Cited on this surface: the Responses API exposes only the URL annotations the answer used, not the wider result pool, so the two columns cannot differ here.

4 of 12 mentioned · 2 ranked · 3 cited · 3 retrieved

Unbranded prompts: 0 of 8 mentioned · 0 ranked · 0 cited · 0 retrieved. Entity prompts: 4 of 4 mentioned · 2 ranked · 3 cited · 3 retrieved.

Claude — claude-haiku-4-5-20251001

Retrieved comes from the web_search_tool_result blocks (every result the search returned); Cited from the answer's citations.

3 of 12 mentioned · 0 ranked · 0 cited · 0 retrieved

Unbranded prompts: 0 of 8 mentioned · 0 ranked · 0 cited · 0 retrieved. Entity prompts: 3 of 4 mentioned · 0 ranked · 0 cited · 0 retrieved.

Gemini — gemini-3.6-flash

Retrieved comes from groundingChunks (the pages grounding fetched); Cited from groundingSupports (the chunks the answer was attributed to).

3 of 12 mentioned · 0 ranked · 3 cited · 3 retrieved

Unbranded prompts: 0 of 8 mentioned · 0 ranked · 0 cited · 0 retrieved. Entity prompts: 3 of 4 mentioned · 0 ranked · 3 cited · 3 retrieved.

Perplexity — sonar

Retrieved comes from search_results plus citations; Cited from the citations array alone.

5 of 12 mentioned · 2 ranked · 4 cited · 4 retrieved

Unbranded prompts: 1 of 8 mentioned · 1 ranked · 1 cited · 1 retrieved. Entity prompts: 4 of 4 mentioned · 1 ranked · 3 cited · 3 retrieved.

Run history

Mentions per surface, run by run, out of the prompts that were actually answered — a provider outage shows as errored, not as zero mentions. Runs are appended and never overwritten; a run is one pass over every prompt on every surface, and a run that covered only some surfaces is listed as partial here and never drives the tables above. Answers vary between runs even with nothing changed, so read the direction over several runs, not the difference between two.

Who the assistants do cite and name

Across the unbranded prompts in the latest complete run: the domains the answers cited most, and the products they named most. This is the honest picture of the category as the assistants currently describe it, direct competitors included. A count is the number of answers (out of the unbranded prompts on every surface whose search fired in that run) that cited the domain or named the product at least once; a surface whose search did not fire contributes nothing here. Product names are merged when they differ only in punctuation or an "AI" suffix.

Sources cited most

Products named most

Entity accuracy

For the entity prompts we also check one thing in the answer text: whether it names at least one of the founder's current ventures. Hilyt is one of them, so on a prompt that names Hilyt the check passes whenever the answer repeats the name — it is most informative on the founder prompt. That is a keyword test, not a fact-check — an answer can name a current venture and still describe an outdated identity alongside it, and the check cannot tell. The answer text itself is not published, because it can quote third-party pages; the test runs inside the database and only the yes/no leaves it.

• OpenAI — names a current venture: yes; mentioned: yes; cited: no
• OpenAI — names a current venture: yes; mentioned: yes; cited: yes
• OpenAI — names a current venture: yes; mentioned: yes; cited: yes
• OpenAI — names a current venture: yes; mentioned: yes; cited: yes
• Claude — names a current venture: no; mentioned: no; cited: no
• Claude — names a current venture: yes; mentioned: yes; cited: no
• Claude — names a current venture: yes; mentioned: yes; cited: no
• Claude — names a current venture: yes; mentioned: yes; cited: no
• Gemini — names a current venture: yes; mentioned: no; cited: no
• Gemini — names a current venture: yes; mentioned: yes; cited: yes
• Gemini — names a current venture: yes; mentioned: yes; cited: yes
• Gemini — names a current venture: yes; mentioned: yes; cited: yes
• Perplexity — names a current venture: yes; mentioned: yes; cited: yes
• Perplexity — names a current venture: yes; mentioned: yes; cited: yes
• Perplexity — names a current venture: yes; mentioned: yes; cited: yes
• Perplexity — names a current venture: yes; mentioned: yes; cited: no

Limitations

• Cadence is on demand. Runs happen when we trigger them, not on a schedule, and the first run was 2026-09-02. Until several runs exist this page is a snapshot, not a trend — with one run, n is 1 and any difference between surfaces is as likely to be noise as signal.
• There are no confidence intervals yet. Each cell is a single answer; the same prompt re-asked a minute later can list different products. Intervals will be added once each prompt has enough repeats to compute one honestly.
• These are API surfaces, not the consumer products. The ChatGPT, Claude, Gemini and Perplexity apps add their own instructions, memory and interface, and what they show a user can differ from what the same vendor's API returned here.
• The prompt set is the fleet benchmark's twelve prompts, chosen by us and written in buying-intent phrasing. Every prompt is wrapped in the same fixed instruction to search, answer normally, and finish with a ranked TOP line of recommendations — the list is requested, not volunteered, and a model that would never offer a ranking unprompted still produces one here. Category and problem prompts are unbranded; the entity prompts are not, and they are reported separately for that reason. Other phrasings, and no wrapper, would get other answers.
• Mentioned, ranked and cited are detected by matching Hilyt's names and domains in the returned text, list and citations. A mention under a name we do not match would be counted as a miss. Retrieved is only as wide as each API exposes: Claude's search-result blocks and Perplexity's search_results list a pool wider than what was cited, Gemini's grounding chunks do too, but OpenAI's Responses API exposes only the sources the answer used, so on that surface Retrieved equals Cited by construction.
• Product names in the tables are the assistants' own words, lightly merged. They are not verified to exist, and a name appearing here is not an assessment of the product.
• We run this benchmark ourselves, about ourselves. The prompts, the detection rules and the raw observations are ours; the dataset is published so the counts can be checked, but there is no independent auditor.
• This is the visibility measurement plan-33 separates from technical readiness. The AI readiness check asks whether a page can be fetched and parsed; this page asks what assistants return. How the readiness score works describes the other measurement.

How the AI readiness score works · Run the AI readiness check · Dataset (JSON) · Guides · hilyt.it