How we grade

The grade describes the claim, never the finding

This is the rule that took us longest to get right, and we got it wrong first. A grade always attaches to the claim as it circulates. It never attaches to the study that settles it.

Take "you can track your brand's ranking position inside AI answers". SparkToro's research demolishes it: asked the same question a hundred times, ChatGPT or Google's AI returns the same list of brands fewer than once. We verified that study against the primary source. Our first draft therefore graded the claim VERIFIED, which read as though we endorsed the very thing the study destroys.

The claim is REFUTED. The study is what refutes it. Confusing the two is how a citation list turns into an endorsement, and it is logged in corrections.

What we do before publishing a number

1. Find the primary source. Not the blog citing the blog citing the study. The study.
2. Fetch it and read the number off it ourselves. If we cannot surface the figure from the source, the claim is UNREPRODUCED, even when we believe it.
3. Record the sample size, the date and the method. A number without an n is an opinion wearing a decimal point.
4. Check what the source actually measured. The most-quoted statistic in this field measures a research benchmark, and is quoted as though it measured ChatGPT. That gap is where MISLEADING lives.
5. Declare the publisher's interest. If a finding supports the value of a product its publisher sells, it carries the VENDOR-INTERESTED flag. Ahrefs, Semrush and Profound all publish genuinely useful, large-sample research. They also all sell something. Both facts get printed.
6. Say who checked, and when. Every claim carries a checked date. Evidence rots.

What VENDOR-INTERESTED does not mean

It does not mean wrong. The best datasets in AI search belong to companies with a commercial stake in AI search mattering: Ahrefs, Semrush, Profound, Peec. Refusing to cite them would leave us with almost nothing and would be its own kind of dishonesty. The flag exists so you can weigh the incentive, not so we can dismiss the work.

A vendor study with a disclosed sample size and method beats an independent blog post with neither. That is not a compromise, it is the whole point of grading rather than tribalism.

How we treat our own conflicts

We publish the exact affiliate terms of every tool we review, including the ones that pay us nothing, on who pays us. There are currently 5 tools on this site that cannot pay us a penny, and they are named. Nightwatch, which pays more than anything else here, is named too, at the top, rather than buried where the incentive is harder to see.

When we are wrong

Corrections are permanent, dated, and kept on their own page. We do not silently edit. If a grade changes, the old grade and the reason stay visible. The first two entries in that log are both our own errors, made while building this site, before a single reader arrived.

If you can show us a primary source we missed, send it. Regrading in public costs us nothing except the pretence of never having been wrong, which was never worth much.

What we cannot do

We do not have Ahrefs' crawler or Profound's citation panel. We cannot run a 4-million-URL study, and any site our size claiming to has just told you something useful about itself. What we can do is read the studies that exist, check whether their numbers survive contact with their own sources, and write down what we find. That is a smaller job. It is also, at present, one nobody else is doing.

There is a harder limit than resources. Two of the largest AI engines contractually forbid the measurement itself: Google's grounding terms prohibit "using Links to build an index", and Microsoft's prohibit "creating a database of Output". So where we measure citations, we measure Perplexity, OpenAI and Anthropic, and our coverage understates Google, the largest AI answer surface there is. We publish that limitation rather than bury it. Both clauses are quoted in full, with their effective dates, on the grounding clause.