Read, Used, and Never Named: The Third Way Your Firm Goes Missing in AI Search
On August 27, an analyst published a teardown of a single live ChatGPT conversation. The question was ordinary and commercial: which live chat support software is best. ChatGPT pulled 221 results to answer it. Eighty-four of those were Reddit threads. It issued eleven citations. Reddit received zero.
Not zero because the threads were unreachable — they were fetched. Not zero because they were irrelevant — the answer reads like someone who had spent an afternoon in those threads. Eighty-four pages entered the machine, shaped what came out, and left no trace on the visible answer.
That is a third state, and most firms have never considered it. It is not being absent. It is not being cited. It is being read and going unnamed.
Three states, not one
When a managing partner says "we don't show up in ChatGPT," they are describing a symptom that has three distinct causes.
Absent. Your page never entered the retrieval pool. The engine did not read you because it never had you. This is a plumbing problem — JavaScript-gated text, unindexed pages, a consultation wall in front of the answer.
Retrieved but silent. Your page was fetched. It informed the answer. Another site got the credit. This is the state the August 27 teardown captured at scale, and it is invisible from every dashboard a small firm has access to.
Cited. Your firm is named and linked. You won — for that query, on that day.
These look identical from the outside. Nobody calls. The consultations do not come. And the fix for each one is completely different, which is why firms spend a quarter rewriting content when the problem was a crawler, or buying technical SEO when the problem was a paragraph that could not be quoted.
The scale of the silent middle
Ahrefs analyzed 1.4 million ChatGPT prompts and found the model cites roughly half the URLs it retrieves — 49.98%, to be precise about an imprecise thing. Half of everything the engine reads gets no attribution at all.
Where a page comes from matters enormously. Broken out by retrieval channel, pages arriving through general web search are cited 88.46% of the time. Reddit content, pulled in through its own dedicated feed, is cited 1.93% of the time — and accounts for 67.8% of all non-cited URLs. The model learns from the crowd and credits the institution.
That last sentence is the opportunity, not the problem. Your firm's site is the institution. If ChatGPT is looking for a page it can put its name behind on adjustment-of-status timelines or RFE response windows, a law firm's own explanation is exactly the kind of source it prefers to cite. The question is whether your page is written in a form it can lift.
One more finding from the same study, and it is the practical one: what separates cited from non-cited pages is how closely the page title matches the engine's internal sub-questions — its fan-out queries — not the user's original prompt. Cited pages scored 0.656 on that similarity measure against fan-out queries. Non-cited pages scored 0.484 against the prompt. Pages with plain-language URL slugs were cited 89.78% of the time versus 81.11% for those without.
Why this is not the ghost-citation problem
In early August we covered ghost citations — the case where your firm is linked in the sources sidebar but never named in the answer text a client actually reads. That is real, and it is a distinct failure.
This is a layer below it. In the ghost-citation case, the link exists; a determined reader could find you. In the silent-retrieval case, there is nothing to find. Your page shaped the sentence and left no evidence it was ever there. The only way to detect it is to ask the engine directly what it retrieved, and then read the answer next to your own page and recognize your own reasoning coming back in someone else's voice.
Knowing which one you are in changes the fix entirely. That is the whole argument for diagnosing before spending.
Do this in the next twenty minutes
- Pick a real question. Take one a client asked you in the last month, in their words. Not "asylum" — "can I still work while my asylum case is pending in 2026."
- Run it in ChatGPT with web search on. Fresh conversation, no prior context, paste it verbatim.
- Ask for the retrieval list. Second message: list every URL you retrieved, including ones you did not cite. This is the step almost nobody takes, and it is where the diagnosis lives.
- Place yourself. Not in the list at all — Absent. In the list, not in the answer — Retrieved but silent. Named in the answer — Cited.
- Apply the one fix that matches. Absent: publish a crawlable, plain-text answer with a natural-language slug, no gating. Silent: rewrite the key claim as one self-contained, attributable sentence that survives being lifted out of the page. Cited: log it, date it, and re-check in four to five weeks.
- Repeat across three more practice areas before you conclude anything about the firm as a whole.
The bigger picture
The same analyst who found 84 uncredited Reddit threads also found the opposite four days earlier: a different query where ChatGPT scoped its search to a single subreddit by name and handed it six of eight citations. Same account, same week, both commercial questions. Whatever governs crediting varies by query, and nobody outside OpenAI can say why.
That uncertainty is not a reason to wait. It is the reason to measure. A platform that decides citations query by query, without publishing its reasoning, cannot be reverse-engineered from a blog post — but it can be observed, one real client question at a time, by any firm willing to spend twenty minutes looking.
The firms that will be visible in 2027 are not the ones with the best theory of how AI search works. They are the ones who know, this month, which of the three states they are actually in.
Clientory helps immigration law firms become visible to the families and individuals who need them most — turning AI search into consultations, and consultations into clients.
See Your AI Visibility Score →