← All posts

44 of 73 scans scored 0. So we dropped the score.

October 8, 2026

Until 2 October, Foreground gave every business one number: an AI visibility score from 0 to 100.

On 28 September we looked at the last 30 days of real scans. 44 of 73 scored 0. The average was 14. ASML, TSMC, Apple and national newspapers scored between 57 and 100. Every small business scored between 0 and 12.

That's not a measurement. That's a size check.

Why the old scan couldn't work

It asked broad category questions. Only big brands win those.

One real scan, a small boutique hotel in Chiang Mai. Question one: "I need a good hotel option — what do you recommend?" No city in it. ChatGPT and Claude asked "where?". Gemini recommended Rosewood Hong Kong. Question two: "Top hotel companies in Chiang Mai." Every AI listed chains: Marriott, Four Seasons, Shangri-La. Only the third question had enough detail for a small hotel to be in the race at all.

So a 0 said nothing about the business. The owner couldn't move it. And we could never show anyone that a fix had worked, because no fix could.

The test, before any code

Nobody asks AI for "a hotel." They ask for a quiet boutique hotel for a couple, in the Old City. A driving school in Rotterdam with automatic cars. An estate agent in Haarlem to sell their house.

So we stopped building and wrote down 11 of those customer niches: 6 in Chiang Mai hospitality, 5 Dutch local services, asked in Dutch. Each niche got 3 phrasings: a short search, a request with details, and "what should I look for, and which ones do that well?". All went to ChatGPT, Gemini, Claude and Grok with web search on, through the app's own code. 132 questions a run, two runs. The pass rules were written down before the first one.

What came back:

  • Small businesses win narrow questions. In 10 of 11 niches, at least 3 of the top 5 named businesses were independents. In 7, all 5 were.
  • Narrow niches have a leader, and it holds. Tamarind Village was named in 11 of 12 answers, by all 4 AIs, in all 3 phrasings. Over both runs: 21 of 23.
  • Broad niches have no leader. "Family hotel with a pool, budget, near the Old City" named 25 different hotels. The most-named got 3 of 12. Nobody owns that one yet.
  • The shortlist is small. Utrecht has around 55 botox clinics. The AIs named 12.
  • Where AI looks depends on the market. In the Dutch niches, the winners' own websites were the main source, plus comparison sites like funda.nl and trustoo.nl. In Chiang Mai: Booking, Agoda, TripAdvisor, Time Out, the Michelin Guide and travel blogs.

One niche failed outright: botox clinics in Utrecht, where multi-location groups held the whole top 5. We wrote that down and dropped it.

Being on the pages AI reads

Then the question everything else depends on: does it matter which pages an AI reads?

ChatGPT, Claude and Grok show every page their search returned, not just the ones they cite. So we took the pages one AI read, and checked who the other three AIs named. In 5 of 6 niches, the quarter of businesses on the most of those pages got named 2.4 to 4.7 times as often as the quarter on the fewest.

That's a correlation, not proof. Popular businesses get listed and named. But one case showed the mechanism. The same small hotel from above had been named in 0 of 47 answers. Then ChatGPT named it twice. Both times it cited the same single review page, on a Thailand travel blog that ChatGPT reads more than any other site for that niche.

One page the AI reads. One mention. That's the whole idea behind the fix kit.

The number moves. How much?

Same questions, asked twice, a few days apart. For 64% of businesses the count moved by at most 1 of 12. For 90%, at most 2. For 97%, at most 3.

So that's the rule in the app. Between two scans, only a change of 3 or more of 12 counts as real. And "you improved" has to hold across 3 scans, not one lucky jump. Slower. True.

The phone apps aren't the API

Marc asked the same questions in the ChatGPT and Gemini phone apps. The apps named businesses our measurement never did. One Chiang Mai hotel came up in both apps, and in 0 of 75 API answers.

The main reason: for local questions, the Gemini app answers from Google Maps, not web pages. When we asked Gemini with its own Google Maps tool, most of those app-only names showed up.

So local niches now get a second count, "on Maps: X of 3", shown separately from the web count. And we say plainly what we measure: what AI answers when it searches the web, plus Maps for local businesses. Never "what your ChatGPT app shows".

Writing the fix, then checking it

A count alone doesn't help an owner. So before building the fix kit into the app, we made 6 by hand, for real businesses, from their own websites only.

We planted 5 false claims in each kit. The fact check caught all 30. The rules: every number in the kit has to appear in a source. "Best" or "award-winning" needs a source that says so. A claim that fails is removed, not reworded.

The surprise was what the kits found, not what they wrote:

  • A hotel's best-known website says, on the page, that it isn't the hotel's official site. It's an affiliate.
  • A riverside restaurant's site never says where it is, or that it's by the river.
  • An estate agent's own pages show two different price tables.
  • An estate agent doesn't sell houses at all, the one thing its niche asked for.

So every kit now puts problems like these at the top, tagged "first". And the fit check moved to the start: Foreground reads your site before you pick a niche, so you don't spend a scan on one your site can't support.

Marc read two of the kits. His verdict: "okay, I guess." That's not a glowing review. It's also not the real test. The real test is a business owner reading one.

What we don't know yet

Whether the fix kit moves the number.

We're testing it on three of Marc's own sites, Foreground included, with a test niche and an untouched control niche each. The first two measurements: all six niches, 0 of 12. Not named, not read, not cited anywhere. So a mention later will be easy to spot.

Until that shows something, we don't claim the kit works. We claim what we measure.

What changed

Before: "You scored 0."

Now: "For couples looking for a quiet boutique hotel in the Old City, you're named in 0 of 12 answers. Tamarind Village is named in 11. Here are the pages AI read to answer. You're on none of them. Here's what to fix first."

The same zero. One you can do something about.

— Marc. Run a free scan: one niche, 12 answers, no credit card.