Conformance is not citation
- A perfect agent readability score says software can read and use your site. It says nothing about whether a model recommends you.
- Conformance is measured on your own property. Citation is measured on someone else's, which is why you cannot fix it by editing your site alone.
- Run them as two numbers. A firm that conforms and is never named has an evidence problem off site, not a markup problem.
- The pairing is the finding. Same score, different outcomes, and the gap between them tells you where the work is.
- Anyone selling one number for both is selling you the half that is easier to measure.
Two posts ago I said the browser had started grading sites for agents. Last time I published the rubric I use to grade them more thoroughly. Now the part that keeps the whole thing honest: a firm can pass every check on that list and still never come up when a buyer asks an AI system who to hire.
I have seen it. Clean markup, content in the response, sitemap exact, files published, and the model names four competitors and not them. Nothing on the site was wrong. The site was never the problem.
The takeaway
Conformance is measured on property you own. Citation is measured on property you do not. Run both, and the distance between them is the actual work order.
Why the two numbers come apart
Conformance asks a closed question. Does the file exist, does the text arrive without a browser, do the facts agree. Every input is on your own domain, which is why it is scoreable at all, and why a script can settle it without asking anyone.
Citation asks an open one. When somebody describes your category to a model, does your name come out, and is what follows it true. The inputs there are mostly other people's property: trade coverage, directories, regulator filings, forums, reviews, the pages a system happens to retrieve at the moment it is asked. You can influence that. You cannot edit it, and no change to your own HTML reaches it directly.
Google says as much in its own way. Its guidance on AI features tells you that no special file, markup or writing style gets you into generative results. Read that alongside the Agentic Browsing audit and the division is clear enough. One measures whether the machinery works. Neither promises the outcome.
The four cases
Put the two numbers on the same page for a firm and you land in one of four positions, and each one has a different bill attached.
Conformance low, citation low. The common case, and the cheapest to start on, because the site work is defined and finite and you own all of it.
Conformance high, citation low. The frustrating one, and the one that gets misdiagnosed most. The site is fine. The evidence about the company that lives elsewhere is thin, stale or contradictory, and the fix is publishing, correcting records and earning coverage, none of which is a web project.
Conformance low, citation high. Reputation is carrying a site that cannot be read. This is the fragile position: the models are describing you from other people's pages, so the description drifts and you have no way to correct it, because the correction lives on a site nothing can parse.
Both high. Then stop buying visibility work. If you are read accurately, named accurately, and still losing, the problem moved downstream to the offer or the proof, and I would rather say so than keep selling.
How to measure the second one without a dashboard
Fix a set of questions in the language a buyer would actually use, ask them across the systems your buyers use, and record three things each time: whether you are named, whether the description is accurate, and what is cited underneath it. Keep the question set fixed so runs are comparable. That is the whole method. The rigor lives in the fixed questions and the record, not in the tool.
I have written the longer versions of this elsewhere: how to measure AI visibility on why a single score collapses six different observations, the query library on keeping the questions fixed so the runs mean something, and the grounding gate on the difference between a model mentioning you and a model citing a URL it actually read.
What I would do with the pair
Score conformance first because it is cheap, mechanical and entirely yours. Then run the citation read against a real competitor field, not against the market in the abstract, since “are we visible” has no answer and “are we visible against these five firms” does.
Then look at the gap and let it pick the work. A wide gap in the conformance direction is a build. A wide gap the other way is a publishing and evidence problem wearing a website costume. Same two numbers, opposite invoices, and you cannot tell which one you are looking at with only one of them in hand.
Common questions
If my site is fully machine readable, will AI systems recommend me?
No. Readability is a precondition, not a cause. It means the software that arrives can parse what it finds. Whether a model names you in an answer depends mostly on what exists about you elsewhere: coverage, directories, filings, reviews, the places a model was trained on and the places it retrieves from now. You control your own site completely and that other evidence barely at all, which is exactly why the two have to be measured separately.
How do I measure citation without buying a dashboard?
Write down the questions a buyer would actually ask, in their words, then ask them across the systems your buyers use and record what comes back: whether you are named, whether the description is accurate, what is cited as the basis. Do it on a fixed set of questions so the runs are comparable over time. The discipline is in the fixed question set, not in the tooling.
What if I score well on both and still lose deals?
Then the problem is not visibility, and I would rather tell you that than sell you a program. Being found, described accurately and recommended is the top of the funnel. If a buyer reaches you accurately informed and still goes elsewhere, the issue is the offer, the proof, or the price, and no amount of structured data will move it.
Sources
The citation half is the head to head read: how AI systems describe, cite and recommend you against a real competitor field, with the questions written down.