How to tell if the advice you’re getting is real. Including ours.
Most of what’s sold in this category can’t be checked. We ask three questions first: Who measured it? On how many? When?
If a claim can’t answer all three, it doesn’t go in the product. That rule has cost us three statistics we liked.
“5x better”
“AI-referred visitors convert 5x better than organic traffic.”
Removed- Who measured it?
- At least five different “independent” studies, with contradictory dates.
- Fails
- On how many?
- Contradictory sample sizes across those studies.
- Fails
- When?
- One of them is dated before the reports that supposedly corroborate it.
- Fails
It can’t answer all three, so it doesn’t go in the product.It answers what you asked. The rule is all three.
3 of 3 failOur log: There is no traceable primary study. It should never have been used.
“61% of clicks”
“Position #1 on Google loses 61% of clicks when AI Overview appears.”
Removed- Who measured it?
- Seer Interactive. Real and traceable, but a single vendor’s figure.
- Weak
- On how many?
- No sample size in the claim as we printed it.
- Fails
- When?
- September 2025. Now over a year old.
- Weak
It can’t answer all three, so it doesn’t go in the product.It answers what you asked. The rule is all three.
1 of 3 failOur log: Real and traceable, but a single vendor’s figure, and now over a year old.
CHANGELOG.txt, Module 1: the three statistics on the welcome screen.
Things you can safely not do.
Don’t buy an llms.txt file.
- Who says so
- Google, on its own page about generative AI in Search.
- Named there as
- llms.txt and special markup
- In our own product
- Moved from a Day-1 task to an optional extra. The evidence was one study: SE Ranking, nearly 300,000 domains, November 2025.Checked again against Google’s page on 26 July 2026: already described as optional.
- What it covers
- Google Search.
- What it doesn’t settle
- How Perplexity, Claude or Copilot weigh anything.
Google — primary source. Our own CHANGELOG.txt for the rest.
Don’t chunk your content for AI.
- Who says so
- Google, on its own page about generative AI in Search.
- Named there as
- content chunking
- What it covers
- Google Search.
- What it doesn’t settle
- How Perplexity, Claude or Copilot weigh anything.
Google — primary source.
Don’t rewrite your pages “for the models.”
- Who says so
- Google, on its own page about generative AI in Search.
- Named there as
- rewriting content for AI
- What it covers
- Google Search.
- What it doesn’t settle
- How Perplexity, Claude or Copilot weigh anything.
Google — primary source.
Don’t pay anyone for mentions.
- Who says so
- Google, on its own page about generative AI in Search.
- Named there as
- seeking inauthentic mentions
- A second source
- Muck Rack, across 25M+ links in ChatGPT, Claude and Gemini: where the citations came from.
Google — primary source.
Google, May 2026: many of these hacks aren’t effective or supported by how Google Search actually works.
That page is about Google — it doesn’t settle ChatGPT, Perplexity, Claude or Copilot. Where nobody knows, we say so.
Three of our own numbers turned out to be wrong.
Three numbers we used to print failed our own test: who measured it, on how many, and when. If we’ll do that to our own claims, you can guess how we feel about the ones nobody checks.
Three statistics on Module 1’s opening screenRemoved: “5x better”, “61% of clicks” and “98.8%”. The 5x has no primary study behind it at all.
Two things nobody can tell you, including us.
The two largest studies on what actually drives AI citations disagree with each other.
Ahrefs, 75,000 brands
Puts referring domains near the bottom of what correlates, against its own product’s interest.
“all factors revealed moderate to very weak correlations.”
What correlates with being cited · 0 to 1
SE Ranking, 129,000 domains
Puts referring domains at the top.
We only hold its ranking, not its figures, so we draw no bars for it.
What predicts being cited
- 1Referring domainsthe one they disagree onstrongest
Both are real studies. We don’t know which is right. Nor does anyone quoting one without the other.