AI BEACON · PROMPT DIAGNOSIS
Open one prompt and see
what the answer was built from.
Knowing you lost a prompt is not much help on its own. What helps is the list of sources the answer used, whether anything of yours was among them, and what the answer said about you. From there the next move is usually obvious.
A prompt was lost. The fix was a guess.
Without the sources in front of you, remedial work on AI visibility turns into trial and error.
The page was rewritten, then rewritten again, because nobody could tell whether the problem was the content, the crawl or the competition.
The page holds position one for the keyword and the generated answer above it does not mention the brand at all.
One engine recommends you and another does not, which rules out most of the explanations people reach for first.
Four questions, asked of one answer.
Diagnosis is a narrow exercise. You are looking at one prompt, on one engine, and working out what fed it.
Choose the prompt
Usually one you are absent from, or one you have recently lost.
See prompt coverage →Every source the answer used
In full, with cited and consulted sources marked apart.
Look for your own pages
Whether a page of yours was quoted, read without being credited, or never read.
Pick the work that follows
A missing page, a page that cannot be read, or a source worth earning a mention on.
The anatomy of a single answer.
Four things, which between them account for most reasons a brand is absent.
Everything the answer drew on.
Listed in full rather than summarised, because the one that matters is often the fourth or fifth entry rather than the first.
- Resolved to the pageA URL, a named thread or a video, with the domain alongside.
- Cited and consulted apartQuoted sources and sources the model merely read are marked separately.
- Yours markedIf one of your pages appears anywhere in the list, it is flagged.
Read, quoted, or never opened.
These three outcomes lead to completely different work, and telling them apart is most of what diagnosis is for.
- QuotedThe answer credited your page, which is the outcome you want.
- Read without creditThe model consulted your page and chose something else to cite, which usually points at the content.
- Never readNothing of yours was opened, which usually points at something technical.
What the answer said about you.
On a prompt where you are named but passed over, the description is usually the explanation, and it is right there in the text.
- The sentence itselfHow your brand was described in that answer, if it was.
- With the sentiment scoreA reading from 0 to 100 for that prompt.
- Beside the brands namedWho was recommended, and in what terms.
Diagnosis, then the work.
Each outcome points somewhere specific in the rest of the platform, reading the same project data.
- No page covers the questionContent Studio writes one, in your brand voice.
- The page exists and is skippedTechnical Suite applies the technical fixes on your live site.
- A source keeps appearingThat is a mention and a link worth going after.
What this can see, and what it cannot.
Worth being plain about, because the question people want answered is what the model was thinking, and that is not available to anyone.
A typical stack
- –Explanations assembled from a screenshot and a hunch
- –The content rewritten before the crawl is checked
- –One engine treated as all of them
- –No record of what the answer was built from
With prompt diagnosis
- ✓Every source the answer used, listed in full
- ✓Your own pages marked as quoted, read or unread
- ✓The description and the brands named, kept with it
- ✓A clear handover to the page or the fix that follows
This reads what an answer used and what it said. No tool can see inside a model's reasoning, and nothing here should be read as doing that.
The rest of AI Beacon, and where this one sits.
Diagnosis starts from a prompt and ends in a piece of work. These are the pages at both ends of that.
What is prompt diagnosis?
Prompt diagnosis is the examination of a single AI answer: every source it drew on, whether any of your pages were among them, and what it said about your brand. In AI Beacon you open one tracked prompt on one of 7 AI engines plus Google AI Overview, AI Mode and Search, and get the full source list with cited and consulted marked apart, your own pages flagged as quoted, read or never read, and the answer's description of your brand with its sentiment score.
Key facts
- It works on one prompt and one engine at a time
- Every source is listed in full and resolved to the page, thread or video
- Your own pages are marked as quoted, consulted or never read
- Those three outcomes point at content work, technical work or a mention to earn
- The answer's wording and sentiment score are kept alongside
- It reports what an answer used and said; it does not reveal a model's internal reasoning
Last reviewed 15 October 2026 by the SEORCE team.
What this can do, and what it cannot.
No. Nothing can see inside a model's reasoning, and any tool claiming to is guessing. What this shows is evidence: the sources the answer used, where your pages stood in that list, and the words the answer chose. In practice that is enough to pick the right next step.
Position and citation are decided differently. The usual causes are that the page was never read, that it was read and something else answered the question more directly, or that the answer leaned on third-party sources instead. The source list tells you which, and crawlability checks cover the first.
The model opened it and chose not to credit it. That generally points at the content rather than the technical setup: the page may bury the answer, or answer a slightly different question than the prompt asked.
Then start with the technical side. Check whether AI crawlers are allowed through in robots.txt, whether the page carries an indexing directive, and whether the content needs JavaScript to appear. Googlebot renders JavaScript; a Vercel and MERJ crawl study found the AI crawlers it measured were not observed to, so test your own pages.
Yes, on all seven plus Google AI Overview, AI Mode and Search. Doing the same prompt across two engines is often the fastest diagnosis, because the difference in their source lists narrows the cause considerably.
It comes with a direction. A missing page becomes a brief in Content Studio. A page that cannot be read becomes a technical fix, and AutoFix applies fixes of that kind on your live site through the script tag.
Take one prompt apart properly.
Track the questions you ought to win, then open the ones you lose and read what the answer used.