We ran Foreground's own scan on Foreground. Here's what came back.

If you build a tool that checks whether AI models know who you are, the obvious thing to do eventually is point it at yourself. So this week I ran the exact same pipeline a customer gets — real prompts, real models, real web-grounded calls, no mocking — against Foreground itself.

The honest number: 11 of 37 completed checks mentioned us (13 of 50 calls failed outright — Grok and GPT-5 mini's reasoning overhead account for most of that, a known cost of running the full unthrottled research script instead of the hosted app's tuned concurrency). Read as a bare percentage, that's a mediocre score. Read by actually opening the raw responses — the step proof-script's own output insists on before you trust anything it says — it's a much more specific and more useful problem than "mediocre."

Most of the models don't know which Foreground we are. Ask GPT-5 mini who Foreground is and it stops to ask you back: "an AI meeting-notes app, or a brand-visibility tool like Otterly.AI?" Ask Claude Haiku and it finds a real, unrelated Wikipedia-adjacent nonprofit called Foreground AI — "a research collective exploring the possibilities and effects of AI across narrative and culture" — and confidently describes that, correctly, as if it had answered your question. Ask Mistral Small and it describes a "conversational AI platform" with file uploads and SSO that has nothing to do with us. Three different models, three different wrong answers, each delivered with full confidence.

Gemini was the one model that consistently found the real thing — mostly. Its best answer accurately described the 5-prompt, 5-model, web-grounded scoring approach well enough that it read like it had actually visited the site. Its worst answer, on a "what do people say about Foreground vs. Otterly.AI" prompt, invented an entire fictional offering: a "Foreground Digital" agency selling a "$1,997 AI Visibility Foundation Fix" for "Wikidata, schema, and entity graph repair." That service doesn't exist. No one on our team wrote it, priced it, or shipped it. Gemini's web search found something that pattern-matched closely enough to "Foreground" plus "AI visibility" to get stitched into a confident, detailed, entirely wrong paragraph — a textbook hallucination, the kind our own glossary already warns about, now demonstrated on our own name.

Here's the part worth sitting with: our own presence detection counted every one of those as a citation. The regex doesn't know the difference between "Foreground accurately described" and "Foreground confused with a nonprofit" and "Foreground partly invented." It just matches the string. That's not a bug we're quietly going to patch and pretend didn't happen — it's the exact reason how-it-works.html says presence detection is "intentionally conservative" and this repo's own process requires a human to actually read the raw text before trusting a number, every time. We ran that process on ourselves and it caught something real.

Why this is happening to us specifically: "Foreground" is not a rare string. It's a common English word, already the name of a nonprofit, a recorder app, and apparently whatever Gemini's search index decided to merge together. A model with no other signal reaches for whichever "Foreground" it can find the most text about — and right now, that's not consistently us. This is precisely the "off-site authority" gap our own entity_presence check exists to catch (we don't have a Wikipedia page; of course a model reaching for a Wikipedia-shaped answer finds someone else's), and precisely the "specific, quotable claim" gap our clarity_check exists to catch. We're not exempt from the problem we built a product to measure. That's the whole point of running it on ourselves instead of only ever reading the methodology.

What we're actually doing about it: building a clearer, more specific public-facing description of what Foreground is (this post, plus the new glossary, are part of that), and treating our own score the same way we'd tell you to treat yours — a real signal, worth tracking again after we've actually changed something, not a number to be embarrassed about in the meantime.

— Marc. If a search for your own business name turns up a stranger's answer instead of yours, that's exactly what Foreground is built to catch — including, evidently, when the business is us.