You do not need an agency to find out why machines ignore your website. You need 90 minutes, a browser, a notes file, and the discipline to stop when the timer stops. This is the exact audit we run on discovery calls, converted into a solo exercise with the clock built in.
It covers four passes — access, extractability, identity, and a live test against the engines themselves — and ends with a scored checklist plus a punch list ordered by effort. Most owners find eight to twelve fixable problems. Most fixes need an editor, not an engineer.
Set a timer. Ninety minutes, five blocks. Go.
Before the clock: what you need open
Open five tabs: your site, your-domain.com/robots.txt, Google's rich results test, one AI engine with live search (Perplexity is easiest), and a blank notes file with four headings — Access, Extractability, Identity, Engines. Then pick the five pages that make you money: home, one or two service pages, pricing, about. Everything below runs against those five pages, not your whole site.
The five-page constraint is the audit. Trying to assess forty pages in one sitting produces a mood, not a list.
Ninety minutes assumes a site of five to twenty pages. Bigger sites still audit five pages — you just choose them more carefully.
Put your phone in another room. The audit works because it is timed, and the timer only works if you do.
Minutes 0–15: can machines get in?
Load /robots.txt and read it line by line. You are checking whether GPTBot, ClaudeBot, PerplexityBot, or Google-Extended are blocked — a surprising number of sites still deny them because of a 2023-era decision nobody revisited. Then check for /llms.txt, confirm /sitemap.xml exists and lists your five pages, and load your homepage with JavaScript disabled to see what a text-first crawler actually receives.
Score four checks pass or fail: AI crawlers admitted, llms.txt present, sitemap present and current, content readable without JavaScript.
The JavaScript-off check matters because several AI crawlers fetch static HTML and render little or nothing. If your value proposition lives inside a script, some engines have literally never read it.
If the content disappears without JavaScript, note it in red. That single finding reorders your priority list more than anything else in the audit.
Minutes 15–40: the extractability pass
For each of your five pages, read only the headings and the first paragraph under each. Ask two questions: is the heading something a customer would actually type, and do the first 40 to 90 words underneath answer it completely, with no dependence on anything above? Count the sections that pass. Across five pages, most first-timers find fewer than three passing sections out of twenty or more.
Also check each page for one number in the first screen of text — a price, a timeframe, a quantity. Engines quote numbers. Pages without them get summarized instead of cited.
Note the single worst offender: the money page whose opening paragraph could belong to any company in your industry. That page is your first rewrite.
Generic openings are the most common failure in this entire audit, and the cheapest to fix well.
Twenty-five minutes, then move. The rewrites happen later, not today.
Minutes 40–60: identity, authorship, freshness
Now check whether a machine can tell who is talking. On each of the five pages: is there a named human author or owner anywhere? A visible updated date? Does the about page state what you are in one plain sentence — category, customers, place, year? Paste each page into the rich results test and note whether any schema exists and whether it validates without errors.
Then check consistency: does your LinkedIn description match your website description, or do they read like two different companies? Machines resolve identity by matching. Drift costs you.
If the schema step sounds technical, it is paste-and-click. The test tool does the reading — you only note green or red.
A missing about-page sentence fails two audits at once: this one and the engine's.
Score five checks: author present, dates present, canonical sentence exists, schema valid, profiles consistent.
Minutes 60–80: ask the engines
Now the live test. Write five questions a real customer would ask an AI — "best [your service] near [your city]," "how much does [service] cost," "is [your brand] legit" — and run each through Perplexity and one other engine. Record three things per answer: were you mentioned, who was cited instead, and did the engine say anything about you that is wrong. Twenty minutes, ten answers, no editorializing.
Wrong beats absent as a finding. A confident misstatement about your prices or service area means a bad source exists somewhere, and finding it becomes a punch-list item.
Use a private browser window so your own history does not flavor the results.
Resist the urge to fix anything mid-test. You are collecting, not treating.
Save the transcript with a date. This is your baseline, and rerunning it monthly is how you will know whether anything you fix is working.
Minutes 80–90: build the punch list
Turn your notes into one page with three columns: the finding, the fix, and the effort in hours. Order by effort, smallest first. A typical list: unblock AI crawlers (0.5 hours), add llms.txt (1), add dates and author lines (2), write missing meta descriptions (2), rewrite the worst page-top answer-first (3), add FAQ blocks with schema to two pages (3). The ordering matters — early wins fund the will for later ones.
Anything above four hours gets split. "Rewrite all pages" is a wish. "Rewrite the plumbing page top by Friday" is a task.
Do not add strategy items. This list holds only things you watched fail in the last 85 minutes.
Ten hours of listed work is normal. It is also, roughly, the gap between you and the competitor the engines currently prefer.
How to score yourself, and what to fix first
Count your passes across the roughly 20 checks: 16 or more is strong, 10 to 15 means solid bones with real gaps, and under 10 means machines are guessing about you — start with access, since nothing else matters while crawlers are locked out. After access, fix identity before extractability: schema, authors, and dates are hours of work, while rewrites take days.
Rerun the full 90 minutes quarterly. The checks do not change much. Your score should.
Access, identity, extractability, engines. Same order every quarter.
And if you want the graded version with zero effort first, the grader on our homepage runs 22 of these checks in about 40 seconds. The audit above is how you understand the grade. Either door leads to the same punch list.