Of 55 brands scanned on LensCore in September 2026, 39 (70.9%) were named by none of the five AI engines we probe. Only 6 (10.9%) were named by all five. Almost nothing sits in between — AI visibility in this sample is closer to all-or-nothing than to a gradient.
This is every brand that ran a scan on lenscore.app between 1 and 30 September 2026, minus two of our own test subjects. 88 scans, 57 distinct brands, 55 after our own are removed. Each brand was asked category-specific buying questions across five engines, and we counted how often the brand was named in the answers.
We are publishing the script, not just the conclusion. Everything below comes out of scripts/september-visibility-study.mjs in our repository, run against the production database. Re-run it and you get these numbers.
Seven in ten brands were invisible to every engine
| Engines that named the brand | Brands | Share |
|---|---|---|
| 0 — named by none | 39 | 70.9% |
| 1 | 2 | 3.6% |
| 2 | 4 | 7.3% |
| 3 | 2 | 3.6% |
| 4 | 2 | 3.6% |
| 5 — named by all | 6 | 10.9% |
The shape here is the finding. If AI visibility were a gradient, the middle rows would be the fat ones — most brands showing up on two or three engines, a few at the extremes. Instead the two ends hold 45 of 55 brands and the middle holds 10. A brand tends to be either findable by all of them or by none of them.
No engine is generous
| Engine | Brands probed | Named at least once | Share |
|---|---|---|---|
| OpenAI (web search) | 55 | 13 | 23.6% |
| Claude (web search) | 55 | 12 | 21.8% |
| Google Gemini (Search grounding) | 55 | 10 | 18.2% |
| Perplexity (Sonar) | 55 | 10 | 18.2% |
| LensCore Retrieval Engine | 55 | 9 | 16.4% |
The spread between the most and least generous engine is 7.2 percentage points — 13 brands against 9. That is a smaller gap than we expected, and it is the practical takeaway: there is no single engine to optimise for. Chasing one of them is unlikely to move the others, because they are not disagreeing with each other very much in the first place.
What this does not show
This is not a random sample of anything. It is the brands that chose to run an AI-visibility scan. Someone who scans an AI-visibility tool probably suspects they have a problem, so this sample is likely tilted toward invisible brands. The 70.9% is a fact about these 55 brands and not an estimate of any wider population.
We queried each engine once per scan. One run cannot separate a brand that is reliably invisible from one that was unlucky on the day. Nothing here measures volatility or consistency, and we have not claimed it does.
Depth varied. 53 of the 55 brands were measured with 8 questions and 2 with 25. Share of voice is a proportion, so the figures are comparable, but an 8-question result is a coarser instrument than a 25-question one.
Nothing here is causal. We measured what the engines said. We did not run an intervention, so we cannot tell you that doing X would have moved a brand from the 39 into the 6.
Why this post replaced four others
Until today this blog carried four articles citing original research — a 50-brand audit, a 500-page citation study, a measured 18% CTR lift. We checked them against our own database on 8 October 2026 and the studies had not happened. The window that claimed 50 brands held five. No analysis script had ever been committed. We retracted all four.
They were not written to deceive; they were written to a content playbook that rewards original data, and the data was invented to fit the template. That is a worse failure than a wrong number, because it is the kind that spreads quietly. The fix is structural: from now on a claim about measurement ships with the script that produced it, or it does not ship.
Method, in one paragraph
Every completed scan between 2026-09-01 and 2026-10-01 carrying an AI-citation block. Brands de-duplicated by registrable domain, keeping each brand’s deepest scan. Engines grouped by their stable internal source id rather than their display label, because our own retrieval engine was renamed mid-month and grouping by label reports it as two engines. Our own test subjects identified by name match and excluded, with the count reported. Question sets: packs-5x5.v1; benchmark core-25.v1.