AI visibility checklist: twenty things you can actually check
Not tips but checks: for every item you can determine in an afternoon whether it is true of you.
Updated ·
This is the cluster's star point, pulling together what the other recipes cover individually. The editorial rule is that every item must be CHECKABLE — not "have good content" but a statement you can verify about your own site. The order roughly matches the order of repair: access and entity first, because nothing else works without them; then citability; then the discipline of measurement. You do not need all of it at once — with the first third green you have already done more than most. One thing to note: this is not a scorecard. There is no aggregate number to extract from it, deliberately — a composite score hides which layer the problem is in, and the fix differs entirely by layer.
The twenty checks
Your key pages render server-side
Look at the served HTML rather than the rendered view: if the substantive text only assembles after JS, a fetcher sees an empty page.
You parsed robots.txt with a tool, not by eye
Run a parser over every important URL for every AI user-agent — the common error is not a missing rule but an over-broad one.
You know which bot class you allow
Training, retrieval and user-initiated are three separate decisions with three separate commercial consequences.
Your rules point at paths that exist
After a rename the rule stays but matches nothing — and a rule that matches nothing raises nothing.
Your name resolves to one thing
Ask for your name with no context on several surfaces; mixed or foreign results mean entity ambiguity.
You have one stable @id, used everywhere
A different identifier per page does not describe one company many times; it describes many companies once.
Every sameAs URL is live and yours
A sameAs pointing at a deleted item is worse than none: a broken identity claim that reflects on the whole node.
Every page carries the FULL node
A crawler reads each URL in isolation; a stub node IS the entity for whoever fetched that page.
Your five core facts agree everywhere
Name, location, founding, category, size — contradiction costs more than omission.
You know whether you qualify for a registry item
Without two or three independent sources, get those first — an item is a consequence of presence, not a means to it.
You know whether you have a knowledge panel
Check signed out, on another device and region; with no panel, claiming is not an applicable step.
Your sentences assert something when lifted
Read them without surrounding context: if they say nothing alone, they cannot be cited.
Your figures carry a source or a method
A number with no method carries no more weight than an adjective.
Freshness is visible and honest
Advancing a date without changing content buys nothing short-term and costs credibility later.
Visible text says WHO produced it
You can be quoted without being named — an author field in the markup is not enough.
You measure repeatedly, not once
One run is one sample; the variance is the result, and it is the part people throw away.
You ask unaided first
Once you supply the name you measure whether it KNOWS you, not whether you COME TO MIND.
You separate knowing from finding
Ask on a sourced and an unsourced surface: the difference between them is the diagnosis.
You know which layer you are fixing
Retrieval moves in weeks, trained memory in model generations — patience is also an intervention.
You ruled out the usual alternatives
Algorithm update, your own technical change, a new competitor, seasonality — all likelier, because more provable.
Frequently asked questions
- In what order should I work through it?
- Roughly the order given. Access and entity are the foundation: if machines cannot get in, or cannot tell which one you are, content work attaches to an ambiguous entity. Citability follows, and measurement discipline last — though it is worth starting that early so you have a baseline.
- Do I need a developer?
- For roughly half the items, yes — server-side rendering, markup and robots.txt touch code. The other half are content and organisational questions you can handle without one: reconciling your facts, visible authorship, the shape of your sentences, measurement.
- How long does it take to work through?
- CHECKING takes a few hours, given access to your own site and one AI surface. FIXING takes far longer and varies enormously by item: robots.txt is a minute, earning an independent source is months. That is why it pays to walk the whole list and assess before starting anything.
- Is there a total score?
- No, deliberately. A composite number would hide which layer the problem is in, and fixes differ completely by layer — access moves in weeks, trained memory in model generations. A number that averages those two together is noise rather than information.