Citations for AI answers, and the four verdicts
Citation signals are worth about 7% of local-pack weight. The reason this chapter still sits in Part II is that the same records now do a second job entirely.
The two things called "citations"
There is a naming collision in this field, and separating the halves is the most useful move in the chapter:
- A local-SEO citation — a directory listing carrying your NAP. Everything above.
- An AI-answer citation — a source an assistant cites when it answers a question.
They used to be unrelated. They are now connected, because the second is increasingly made of the first.
When an assistant answers "best coffee near here" it does not consult a ranking table; it searches, retrieves pages, and writes an answer grounded in what it retrieved — and for local questions the retrieved pages are overwhelmingly directory and review pages, because that is what the web holds about small businesses.
So a directory listing is no longer only consistency evidence. It is retrievable text about your business inside the corpus the answer is drawn from, which reframes absence:
Not being on a directory the assistant cites is not a lost 7%. It is a document that does not exist. The retrieval layer cannot return what is not there.
How those engines weight what they retrieve is not knowable from outside. What is observable, and cheap to observe, is which domains an answer cited and whether you are on them.
Work the gap, not the list
Generic advice says "build citations" and hands you a directory list. The list is the same for every business in every market, which should be the tell.
The upgrade is a join between two things you can measure on your own business: the domains AI actually cited for your keywords, aggregated over recent live answer checks, and the listings verified to exist for you, the output of Lab 12.1. Work only the gap.
Where this lives in the app. On the AI Visibility page (/b/{businessId}/ai-visibility) those are the Sources cited by AI and Listings consistency cards, with a strip above them showing the overlap: each platform AI cited, how many answers cited it, and your listing's status there.
The Authority card scores the same join as Presence on the sources AI cites — "Listed on N of M directories/platforms AI cites for your keywords" — weighting it 25 out of 100, with listings consistency a further 15. Those weights are a design choice derived from correlation evidence, not a Google fact; Changing the answer sets out what the evidence is and how strong it is not.
Why the citation package fails
The 300-directory blast fails three ways:
- Most of the 300 are scraped aggregator sites no assistant cites.
- They carry data you handed a vendor, wrong the day you move.
- You now own 300 records you will never maintain, each a future inconsistency.
Verify first, fix what contradicts, and add a listing only where an assistant demonstrably looks for businesses like yours and you will keep it current.
What this tool does not do. SEOG verifies citations; it does not create them. There is no button that gets you listed on Yelp. Building a citation means going to the platform, claiming or creating the listing, and passing that platform's own review — by hand, in their dashboard.
Four verdicts, and why "not found" is dangerous
A listing check can honestly return four things, and collapsing them is how tools cause damage:
| Verdict | Meaning |
|---|---|
| Consistent | A listing was located and its fields resolve to your profile. |
| Mismatch | A listing was located; at least one field is a different value. Both values are shown. |
| Not found | No listing was located at all. |
| Can't verify | A listing exists — its address on the platform is known — but its details could not be read. |
The last two must never be merged. "Can't verify" reported as "not found" tells you to create a listing that already exists; follow it and you have made a duplicate.
The same discipline forbids fabrication: a checker that cannot read a listing should say so, not fill the row with your own profile's values and report all-green about listings it never saw.
Why automated checkers report false negatives
Two mechanisms, worth knowing before you trust any tool's citation tab.
Directory pages resist plain fetching. Many block automated readers, or render content a simple fetch never sees, so a checker built only on fetching reports empty for listings plainly visible in a browser.
The AI checker trap. The modern approach asks a search-capable language model to find the listing and report its fields, demanding the answer as JSON so it can be parsed.
That combination silently breaks: instructing a model to reply with only JSON suppresses its web search. It answers from memory, reports nothing found, and your report says "not listed on Yelp" about a business with a Yelp page and 200 reviews. The output looks like data; it is a configuration bug.
The fix is two steps:
- Ask in natural language with search enabled.
- Convert that answer to structured form in a separate call that does no searching.
AI engine probe recipes covers this and its relatives.