My last prompt audited the outputs — what AI assistants actually say about a brand. This one audits the inputs: what your site gives the machines to read. Together they're the loop I run for clients: influence the inputs, monitor the outputs.
The push for this one came from rebuilding my own site. It had real problems I only found by reading it the way an engine does: a title tag that was literally undefined, two conflicting versions of my identity in the structured data, syndicated articles pointing authority at the wrong domain. The design looked fine. The structure was lying.
That's the trap. You review your site by looking at it. An answer engine never sees what you see — it reads source, schema, sitemaps, and dates, then decides whether you're a citable authority or background noise.
The prompt
Paste this into any assistant that can browse (ChatGPT, Claude, Perplexity, Gemini), fill the three brackets, and read the scorecard.
You are an answer-engine readability auditor. Audit [SITE URL] the way
an AI search engine reads it — structure and evidence only, not visual
design.
Context:
- Site owner: [NAME OR BRAND]
- The question this site should be the authoritative answer to:
[e.g. "who is Jane Doe", "what does Acme's product treat"]
- Industry constraints, if any: [e.g. "regulated healthcare — accuracy
and fair balance matter", or "none"]
Inspect the following. Report only what you actually find — quote
evidence, no vibes. If you cannot fetch something, say "could not
verify" rather than guessing.
1. ENTITY. Read the structured data (JSON-LD) on the homepage, the
about/bio page, and one content page. Is there exactly ONE Person or
Organization entity with a stable @id that other pages reference rather
than redefine? Are sameAs links present and pointing at the correct
profiles? Are jobTitle and knowsAbout populated, and do they agree
across pages?
2. METADATA. Check the title tag, meta description, and canonical URL
on the homepage and two content pages. Flag anything undefined, empty,
duplicated (like a name suffix applied twice), or self-canonicalizing
when the content states it was originally published elsewhere.
3. MACHINE SURFACES. Check /llms.txt, /robots.txt, /sitemap.xml, and
whether an RSS feed is declared. Does llms.txt exist and reflect the
site's current content? Do sitemap dates vary by page, or does every
URL carry one identical build date? Is clean text of key content
available (markdown versions, print pages)?
4. FRESHNESS. Find the newest machine-verifiable date on the site — a
dated post, a now/updates page, dateModified in schema. State the date
and where you found it.
5. EXTRACTABILITY. For the question in my context above: is there one
block on this site an engine could lift verbatim as the answer? Quote
the best candidate. If nothing qualifies, say "no extractable answer
exists" and describe what the block should say.
Output:
A) Scorecard — the five areas, each rated strong / weak / missing,
with one line of quoted evidence per rating.
B) Top five fixes, ranked by impact against effort. Name the specific
page and describe what "fixed" looks like.
C) The one sentence you would expect an AI assistant to say about
[NAME OR BRAND] today, based only on what you found.
Reading the results
The scorecard tells you where you stand; part C is the one to sit with. It's the machine's honest summary of you, assembled only from what your site actually provides — not what you meant it to say. If that sentence is out of date, generic, or wrong, the fix list in part B is the path out.
Two cautions from running this on real properties. First, make the model show its evidence — an auditor that paraphrases instead of quoting will happily hallucinate a passing grade. The prompt demands quotes for that reason; hold it to that. Second, in regulated categories the fixes still route through review. A schema change is usually low-risk, but anything that touches claims goes to MLR first, same as always.
The whole audit takes about three minutes. The last site I ran it on had been "recently redesigned" and scored missing on three of five. The design was not the problem.