Most of what this industry sells does not work
This page exists because a project about honest measurement that hid the inconvenient research would be worthless. Everything here argues against the easy version of this idea. It is on the site anyway, near the front, with sources.
Six findings, five of them negative
| Claim | What the data shows | Sample | Verdict |
|---|---|---|---|
| llms.txt improves AI visibility | 97% of llms.txt files received zero requests. AI crawlers do not look for files that do not exist — and largely do not look for the ones that do. | 137,210 domains | REFUTED |
| Conversational SEO methods work | Most methods are “largely ineffective and frequently have a negative impact”. Gains shrink as adoption rises — the game is zero-sum. | C-SEO Bench, NeurIPS 2025 | REFUTED |
| Optimising body text raises citations | 3 of 54 method-domain combinations significant. Optimising body text for citations cut top-20 retrieval by roughly 9%. | survey of 45 studies | REFUTED |
| Hidden prompt injection influences answers | Seven-platform controlled test with hidden instructions: zero followed it. Copilot flagged the page as unsafe. | 7 platforms | REFUTED |
| AI brand measurement is stable | Brand identity explains 1.5% of response variance. Under 1% chance of the same brand list twice in a hundred runs. Reddit lost 86% of ChatGPT citations in four days with nothing about Reddit changing. | 12,933 responses | REFUTED |
| Earned third-party coverage correlates | The one thing that holds up. Branded web mentions r=0.664, YouTube mentions r=0.737. Citation rate rises from 8% on owned domains to 34% via third-party outlets. Correlational — not proof of cause. | 1M+ AI-cited links | SUPPORTED |
Two vendors, one question, answers 20× apart
Both are measuring the share of AI citations coming from social sources. Neither publishes its prompt set. This single chart is the strongest argument on the site for doing measurement properly.
Why this matters
The seeding arithmetic does not close
The largest category of paid AI-visibility service sells seeded community mentions. Here is what the published effect size would actually cost at retail prices.
Five orders of magnitude short
Press releases, not audits
The leading vendor in this space is marketed as “the #1 AI Search Visibility Tool, ranked by independent analysis”. That headline appears verbatim across Barchart, “Sports News Highlights”, “Glamand Fashion News”, “Saintpaul Chronicle” and roughly a dozen marketminute domains. That is a paid newswire syndication footprint, not independent coverage.
Another firm in the category raised at a $1 billion valuation on approximately $6.8M of revenue. Across the entire sector there is not one independently audited case study demonstrating that money moved an AI mention rate.
There is an irony worth naming: the syndication itself is a mention-seeding play, one story pushed across many domains to saturate the index. It may well work better than the product it advertises. Nobody has measured that either.
Because the null result is also worth publishing
If the evidence says paid AI visibility does not work, the reasonable question is why run the experiment. Three reasons.
First, nobody has tested this specific thing. Every study above measures whether an existing brand can raise its citation rate. None measures whether a name already dense in the substrate can acquire a new meaning. Those are different questions and the second is unexplored.
Second, the measurement itself is the contribution. A frozen, hashed prompt set with published raw runs and a pre-registered hypothesis is something this industry does not have anywhere. Producing it is useful even if every result is null.
Third, a published null is a real outcome. If this fails, it fails in public with the data attached, which is more than a billion dollars of category valuation has managed so far. The falsification conditions →