The twenty checks

Two layers. Open yourdomain.com/robots.txt and read it against the AI user agents documented by the major AI companies. Their crawler docs list the names, and the names change, so read the docs rather than a list someone pasted into a blog post last year. Two things to know while you read. The big operators now run more than one crawler each, on the same three-way split: one that collects training data, one that builds a search index, and one that fetches a page live because a person just asked a question about it. A robots.txt line written in a hurry to keep “the AI” out usually blocks all three, and the third one is a buyer’s agent standing on your doorstep. Then ask whoever runs your CDN or firewall what its AI-crawler settings are, because that layer can turn readers away no matter what robots.txt says, and at least one large provider now ships blocking as a default on some plans and on new domains.

Pass: the readers you want inside can get in at both layers, and you can name the person who decided which readers those are.

A default nobody remembers accepting is a fail.

View source on your pricing page, or curl it if you have a terminal handy. You’re looking at the HTML your server sends before any JavaScript runs, because the crawlers that build AI search indexes and collect training data fetch that response and stop; the measurements published so far found no sign that the major ones execute a script, including the scripts they download. The agents that drive a full browser on a person’s behalf do run your code, which is why this check is about the crawlers rather than every reader, and why failing it costs you more in being found than in being bought.

Pass: a stranger could reconstruct your plan table from that raw response alone.

If the page arrives as an empty shell that scripts fill in later, you just found the most expensive rendering decision in the company, and the finding travels: a site that fails this on the pricing page usually fails it everywhere. The fix is a server-rendered or statically built version of every page that carries a claim. That’s an engineering ticket, and it goes to the top of the list.

Published, with a currency and billing terms. Buyer agents carry explicit budget constraints more often than human buyers ever admitted theirs, and to a reader holding one, an absent price is an unanswerable question. “Contact us” can still be a legitimate decision, but check that it is one. If the number is there, confirm it’s in the raw HTML from check 2 rather than painted in afterward by a currency script.

Pass: either a number is on the page, or somebody senior can state the reason it isn’t and has accepted what the absence costs with the machine reader, out loud, recently.

Paste your five most commercially important pages into any AI assistant and ask it to list every checkable claim: anything with a number, a name, a date, or a verifiable fact. Then run the same on your closest competitor’s equivalents.

The pass condition is a comparison: your count isn’t the smaller one.

While you have the assistant open, ask it what questions a buyer would ask that each page leaves unanswered. That part isn’t scored. It will sting anyway.

Pick three load-bearing facts: your price, one plan limit, your flagship feature’s name. Search your own site for each. Fail usually looks like copies that drifted, an old number surviving on a landing page nobody has opened in years because it converted well once.

Pass: each fact has exactly one authoritative page, and everything else points there instead of restating it.

Every statistic on the pages from check 4 says when it was measured, next to the number, in text rather than in a footnote a reader would have to hunt for.

Pass: no undated number.

You have one, it’s public, and its entries carry dates.

Pass: the newest entry is no older than your newest release.

If the last changelog entry predates the last thing you shipped, one of them is lying, and a reader that cross-checks will notice which.

Anywhere you claim uptime or performance, the checkable version should sit within one click: a status page with visible history, a methodology with configuration and date.

Pass: every such claim has its evidence next door.

If you claim nothing of the kind anywhere, you pass by default, though is the silence modesty or a gap? I can’t answer that from here.

Certification logos link to the issuing body or the auditor’s registry, not to a larger version of the logo. Case studies carry a company name, a number, and the customer’s consent to both.

Pass: a stranger could verify one certification and confirm one case study is real without emailing you first.

“Reference available on request” printed beside a named customer reads well to a diligence process; print it only where it’s true.

Your three most-sold features, each called the same thing on the homepage, in the pricing table, and in the docs. A capability renamed for marketing reads to a cross-checking machine as two different claims, one of which it can’t confirm.

Pass: one name per feature.

Spec tables saved as images, prices inside banner graphics, product demos on video with no transcript. A PDF counts as text only if the text in it can be selected; a scanned spec sheet is a picture with a file extension.

Pass: everything you would want quoted exists somewhere as text.

Every listing you technically own on a platform you don’t: G2, Capterra, app stores, marketplaces, partner directories, Crunchbase, the data brokers feeding a hundred smaller listings you’ve never heard of. Most of these were filled in once, during a launch or a funding round, by someone who may no longer work there, and they’ve been speaking for you ever since. Put them in one spreadsheet and check each against your current price and current feature names. This is the afternoon-sized check, and it’s the one I’d least want to skip, because these are contradictions you built yourself, on surfaces a machine has independent reasons to trust.

Pass: the sheet exists, everything on it agrees with your site, and the recurring reconciliation has an owner and a calendar slot before the sheet gets closed.

Find the oldest public complaint about you that never got a reply, wherever it lives. Write the dated, factual, non-defensive answer. Draft it one day, post it another.

Pass: it has a reply, and the reply contains no adjectives about yourselves.

Your top twenty buyer questions, into the major AI assistants, answers and citations logged in a spreadsheet. If no such list exists yet, build it from four kinds of question: the category ask (“best X for a team like ours”), the direct comparison against your nearest rival, the task your product exists to solve phrased the way a buyer would phrase it, and your own brand name asked cold. Three or four assistants is enough: the ones with the largest audiences, plus whichever one your sales team keeps hearing prospects mention. For each answer, log the date, whether you were named, any price it quoted, and every domain it cited. It takes an hour to run and another to read. What you’re looking for: your absence from questions you should own, a stale price quoted with confidence, a competitor’s framing of you, and where the citations point, because that short list of domains is where this channel’s influence lives.

Pass: the log exists, you’ve read it, and it has a next-run date.

AI referrals separated into their own channel in your analytics, with conversion rate by source beside it. Your analytics tool sees an AI referral as a session whose referrer is one of a short list of domains, the chat products and the AI search surfaces, and some of the traffic arrives with no referrer at all, so whatever you count is a floor. Build the channel from that domain list, name it, and put conversion rate next to it. The configuration is roughly an hour.

Pass: you can state this month’s AI-referral session count and its conversion rate from one screen, without anyone exporting anything.

Pick five facts stated in both: a limit, a tier boundary, an integration, whatever your buyers check.

Pass: five matches out of five.

These two surfaces get held side by side in a machine’s working memory more often than any other pair you publish, and a mismatch between them is the cheapest contradiction to find and the most expensive to explain.

Software gets judged on review sites and in forums before anyone reaches your domain.

Pass: your newest G2 or Capterra review is younger than your newest major feature, and you can name the last thread about your category, anywhere, that someone from your company read.

Answering threads is optional and has rules. Reading them costs an hour a month and pays for itself.

Your comparison pages name at least one matchup you lose and say plainly which buyer should pick the other product.

Pass is a yes or no.

The reader building the shortlist has read your competitor’s version of the same page, and if neither of you concedes anything, the two pages have the same shape and it has nothing to choose on.

Pull ten products from your feed and check them against their live pages: price, stock, shipping cost, return terms, all agreeing, and the descriptive attributes filled in past the required ones, because the optional fields are the constraints shopping agents filter on. The attributes worth the typing are the ones a constraint can hang on: material, dimensions, weight, care instructions, compatibility, country of origin.

Pass: ten out of ten agree and the attribute columns aren’t half empty.

If ten fail, the fix is a catalog project, and it will need whoever owns the catalog before it needs a copywriter.

At least one commerce platform has switched on agent access by default on behalf of its merchants. Shopify’s agentic storefronts went live by default for eligible stores in 2026, syndicating catalogs to the chat products and the AI search surfaces through Shopify Catalog, with the settings under Sales channels in the admin and a default option named, without irony, “Allow Shopify to manage for me.” Other platforms have so far shipped the same reach as an app install or an early-access opt-in, which is the version you’d remember agreeing to.

Pass: you can state which agent surfaces can currently reach your catalog, which of them can complete a purchase, and who chose that.

The word “currently” is doing real work, because the answer changes without notice. Two footnotes. Turning catalog access off does not stop a crawler reading your product pages; that’s check 1’s territory. And if the answer is “I don’t know,” the check has done its job.