Chapter 10 of 14

Are We Cited?

Everything so far has been prediction: signals that should make you citable. This chapter stops predicting. You put the question from chapter 1 to four answer engines and read what they say — including the name of whoever they cited instead of you.

The problem

GEO has a measurement problem that SEO never had. In search you could open the results page and see your position; the ranking was its own feedback. An answer engine gives you a paragraph. There is no position, no list, no explanation — just a reply that either mentions you or does not, and no way to know which unless you look.

So people optimise for GEO entirely on faith, for months, with no idea whether any of it worked.

The probe

From a page's report, type the question a user would actually ask — the sentence you wrote down in chapter 1, not the keyword. Sjá puts it to ChatGPT, Claude, Perplexity and Gemini (through Google Search grounding) with your own keys, concurrently, and judges each answer: does it cite your site, at which position among the sources, and who was cited instead.

"what is a good notes app that keeps my notes as plain files I own?"

ChatGPT       not cited    obsidian.md, logseq.com, zettlr.com
Claude        not cited    obsidian.md, joplinapp.org
Perplexity    CITED #3     obsidian.md, logseq.com, freddy.app
Gemini        mentioned, no link

Three facts arrive at once, and the third is the most useful.

You are citable. Perplexity did it, so nothing is structurally disqualifying you.

Gemini mentioned you without linking. That is its own category in Sjá, kept out of both the cited and the not-cited counts — because a mention is neither. Counting it as a citation flatters you; counting it as an absence hides that the engine knows you exist.

Obsidian is in every answer. The competitor list is the single most actionable line the probe produces. Open those pages. They are not better written than yours; they are more thoroughly resolvable — entity, dates, links, and a first paragraph that answers the question in one self-contained sentence. Everything chapters 4 and 7 said, visible in a competitor.

Two things the interface insists on

A probe is one sample of a non-deterministic system. Ask the same question twice and you may get different sources. What to read is the trend over several probes and the competitor list — never a single answer. Sjá says this in the interface rather than in a footnote, because a single “not cited” has sent people into rewrites they did not need.

These are APIs, close to but not identical with the consumer products, and each probe costs a few cents on your own account. A daily probe on four engines is four API calls a day until you pause it, and the interface says so where you schedule it rather than on your invoice.

Why not, not just whether

“Not cited” used to be a dead end. It hides two completely different situations with opposite fixes:

  • A content problem — the page does not really answer that question. The fix is writing.
  • An authority problem — the page answers it well and the engine chose better-known sources anyway. The fix is elsewhere and slower, and rewriting the page will not touch it.

With a TypeSafe key, every probe is also judged by Jev from the page's own text and the engines' answers, and the card says which one you have:

Not cited    ChatGPT, Claude
Coverage: 0.89   the page answers the question
-> authority problem. Cited instead: obsidian.md, logseq.com

Two more judgements come with it. Fidelity checks each claim a citing engine makes about you against the page — so a citation that gets your price or your platform wrong shows as misrepresented, with the claim quoted, and a scheduled probe raises an alert the first time it happens. Being cited incorrectly is a worse outcome than not being cited, and it is the one nobody was watching for. Stance says whether an answer recommends you, lists you, references you or advises against you — so “mentioned” finally says whether the mention helps.

Where the boundary is. This is opt-in and off until you add the key, and the app states the consequence where you add it: with a key, the probed page's main text and the engines' answers leave your machine for TypeSafe. Without one, probes work exactly as before. Jev's answers are shown and stored and never acted on — nothing rewrites your page because coverage was low — and a stance is only stated from 0.5 confidence, because a confident wrong “advises against” would send you into a panic over a neutral answer.

Schedule it

One probe is an anecdote. Schedule the question daily or weekly and you get the thing that is actually worth having: a trend, and an alert the day an engine starts or stops citing you — naming who took your place.

That alert is the closest thing GEO has to a ranking notification, and it arrives without you remembering to look.

What you learned

  • Stop predicting; ask. Four engines, your own keys, your real question.
  • The competitor list is the most actionable output. Open the pages that beat you and read their first paragraph.
  • Mentioned without a link is its own category, because it is neither a citation nor an absence.
  • A probe is one sample of a non-deterministic system. Read the trend, never a single answer.
  • 'Not cited' splits into a content problem and an authority problem — opposite fixes, and coverage tells you which.
  • A wrong citation is worse than none, and fidelity is what catches it.
Next: in Chapter 11 you close the loop from chapter 6 with evidence rather than inference: your own server logs, and which crawlers actually arrived.