Where I stop

What I will not sell you

Five things you can buy elsewhere in AI search that I will not sell, and the evidence behind each refusal.

AI search is new enough that most of what gets sold in it has never been tested. Numbers get repeated between blog posts until they sound like facts, and by the time they reach a proposal nobody can name where they came from.

I check. When a claim does not survive being traced to a primary source, I do not sell the service built on it, even when clients ask for it by name and even when competitors are billing for it. This page lists the ones that come up most, so you can see the reasoning before you talk to anyone, including me.

An llms.txt file to improve your AI visibility

How it is usually pitched

A small file at the root of your site that is said to tell AI engines how to read it.

Why I will not sell it

Google has stated that no AI system currently uses llms.txt. That is the vendor speaking about its own products, which is the strongest kind of evidence available on this question.

Two independent log studies agree. Across roughly 137,000 domains, 97% of published llms.txt files received zero requests in a month, and AI retrieval crawlers accounted for about 1% of the requests that did happen. A separate analysis across around 300,000 domains found no correlation with being cited, and reported that including the file made their model less accurate rather than more.

The detail that settles it: AI crawlers do not even check whether the file exists. A file nothing requests cannot influence anything.

Instead

I make the pages themselves readable: server-rendered content, clear structure, and answers stated directly. That is what the crawlers actually fetch.

Schema markup as a way to get cited by AI

How it is usually pitched

Add structured data and AI engines will start quoting you.

Why I will not sell it

A difference-in-differences study of 1,885 pages that added JSON-LD, measured against roughly 4,000 control pages, found no meaningful lift in AI citation. For AI Overviews specifically it found a small decline.

Google’s own documentation says there is no special schema.org structured data you need to add for its AI features.

This is a refusal about one specific claim, not about schema. I add schema on almost every project, because it earns rich results and it makes your entities unambiguous to search engines. I will describe it that way rather than attaching it to a citation promise it does not deliver.

Instead

Schema for rich results and entity clarity, priced and explained as exactly that.

A guarantee that ChatGPT will cite you

How it is usually pitched

Guaranteed mentions in AI answers, sometimes with a specific number attached.

Why I will not sell it

Citation is a long tail. The ten most-cited domains account for only around 12% of ChatGPT citations, so the remainder is spread across an enormous number of sources. Nobody controls that distribution.

Results are also unstable between identical queries. Only about half of the entities named in an answer persist when the same question is asked again, and only a small share of citations appear across more than one engine. A guarantee would have to survive that instability, and none can.

The deeper problem is attribution. Even when a citation appears, no agency can honestly prove its own work caused that specific mention rather than a review, a forum thread or a press piece published the same month.

Instead

I measure share of voice across many prompts over time, and I show you the trend without claiming credit for individual mentions.

Statistics about what AI engines reward

How it is usually pitched

Tables get 2.5 times more citations. Lists with four or more items get 67% more.

Why I will not sell it

I tried to trace these two numbers and could not. They lead only to aggregator articles citing each other, with no primary study anywhere in the chain. As far as I can establish, they were invented and then repeated until they sounded established.

The underlying advice is not wrong. Clear structure genuinely does help, and there is real peer-reviewed work supporting answer-first writing, plain structure and citing your sources. What is wrong is the false precision, and false precision is how you tell who checked.

If I quote a number to you, I will name the study and tell you what it actually measured. Several of the honest ones measured word share on a benchmark rather than live traffic, and that distinction matters when you are deciding where to spend.

Instead

I give you the direction the evidence supports, and say plainly where the evidence runs out.

Fake or incentivised reviews, and undisclosed endorsement

How it is usually pitched

Seeded reviews, bought engagement, or paid recommendations that do not look paid.

Why I will not sell it

The FTC rule on consumer reviews and testimonials, in effect since 2024, makes several of these practices directly actionable with civil penalties. This is legal exposure, not a reputational risk you can decide to absorb.

Liability can attach to the brand as well as the creator, so an agency doing it quietly on your behalf does not protect you. It creates a problem with your name on it.

It is also self-defeating in AI search, where third-party corroboration matters precisely because it is independent. Manufacturing it removes the property that made it worth having.

Instead

I build the conditions for genuine reviews and coverage, and I put disclosure in the contract rather than in a polite reminder.

Where this page could be wrong

These refusals reflect the best evidence I can find today, not permanent truths. AI search is moving quickly, and some of this will age.

The llms.txt position is the most likely to change: it is a refusal based on adoption, not on the idea being bad, and if an engine starts using the file I will say so here and change my advice. The schema study is one study, well designed but unreplicated, and one study is thinner evidence than I would like.

If you find a primary source that contradicts anything on this page, send it. I would rather update this page than defend it.

If that is the way you want to be sold to

Tell me what you are trying to fix and I will tell you honestly whether I am the right person for it.

Start a conversation Read the full method