ChatGPT Stopped Searching and Started Looking Things Up

On August 8, ChatGPT's use of the `site:` operator went from about 0.4% of its fan-out queries to nearly 17% — in a single day. That is roughly a 46-fold increase, measured by the AI visibility platform Promptwatch, which tracks the sub-queries ChatGPT writes for itself when it decomposes a prompt into web searches.

The same day, the average number of searches per response nearly doubled, from about 1.08 to 1.83. Those two numbers moving together is the part that matters. If ChatGPT had simply swapped its old open-web queries for domain-scoped ones, the search count would have stayed flat. Instead the `site:`-scoped queries appear to be stacked on top of the broader searches it was already running.

Plain version: ChatGPT didn't stop searching the web. It started also going to specific, named websites and asking them directly. That is a different retrieval behavior, and it changes what your practice has to be in order to appear.

Retrievable was the old bar

For two years the working assumption in AI visibility has been a retrieval problem: get your pages indexed, keep them crawlable, write them so a model can lift a clean answer out of them, and you become eligible to be cited.

A domain-scoped query inserts a step before that. To run `site:yourfirm.com adjustment of status interview timeline`, the engine has to already have decided that yourfirm.com is a place worth asking. That decision happens before any page of yours is fetched. Crawlability doesn't get you into it. Being well-written doesn't get you into it. The engine has to hold your firm — the name, and the domain attached to it — as a known destination for that kind of question.

Retrievable is now the second bar. The first is being named.

What the collapse actually demonstrated

Most of the coverage of this change arrived through Reddit. Reddit's share of ChatGPT Search citations averaged 3.83% from July 18 through August 7, then fell to 0.52% across August 14–17 — an 86.4% relative drop. Google's AI surfaces showed nothing like it over the same window: AI Overviews declined roughly 11%, AI Mode roughly 30%, both gradually.

Two boundaries before anyone builds strategy on this. Promptwatch is explicit that its data shows when the shift happened, not why — the finding is provisional, and the mechanism is inference. And it isn't a blanket demotion. On August 27, Search Engine Journal published a capture of ChatGPT scoping a query to a single subreddit by name, with a 3,650-day freshness window, then handing six of eight citations to threads from it. Reddit is still being asked. It's being asked deliberately, by name, on the questions where a forum is honestly the best source.

That is the useful finding, and it's the opposite of a penalty story. The engine is behaving less like a search box and more like someone who already knows where to look. Discussion platforms lose when the question has an authoritative named home. Official sites, documentation, and institutional domains win — because those are what a domain-scoped query targets.

For an immigration practice, that is a favorable structural change and a demanding one at the same time. Favorable, because your own site is exactly the kind of authoritative destination this behavior rewards. Demanding, because it only helps if the engine already associates your firm with the question being asked.

What makes a practice a named destination

Three things, roughly in order of how much they're neglected.

A stable name-to-domain link. The model has to know that "Ramirez Immigration Law" means `ramirezimmigration.com`. That association is built from consistency across everything the engine has read — your Google Business Profile, your state bar listing, your AILA directory entry, your LinkedIn page, court filings, press mentions, your own footer. A firm that appears as three name variants across four platforms has given the engine three weak associations instead of one strong one.

A topical reason to be asked. Nobody runs a domain-scoped query at random. It happens when a domain is already associated with the subject. If your site has one thin "Immigration Services" page covering H-1B, asylum, naturalization, and removal defense in nine sentences, there is no subject you own. If you have a substantive, dated, attorney-bylined page on what a Request for Evidence on an I-140 typically asks for, you have one.

Pages that survive being asked. A `site:` query can only return what's indexed. Content behind a form, in a PDF that was never indexed, in a JavaScript-rendered tab, or on a page Search Console flags as Crawled – currently not indexed is invisible to a domain-scoped search even when the engine deliberately targets your domain. Thin and unindexed pages now cost you twice.

What to do this week

1. Run a site query on your own firm. In a plain search engine, run `site:yourfirm.com` plus the question a client asks you most. Look at what comes back. If the top results are your homepage and a contact page, that is what a domain-scoped AI query has to work with too.

1. Check your crawler logs for August 8. In your server logs or hosting analytics, compare `OAI-SearchBot`, `GPTBot`, and `ChatGPT-User` activity in the four weeks before August 8 to the four weeks after. If ChatGPT is now asking your site directly, it should show up as crawl activity — earlier and more reliably than a citation count would.

1. Fix the name-to-domain link. Pull up your Google Business Profile, state bar listing, AILA directory entry, LinkedIn page, and Avvo or Justia profiles side by side. Firm name, address, phone, and URL should match character for character. This is the cheapest item on the list and the most commonly wrong.

1. Pick one question and own it. Choose the question you field most in consultations — the RFE question, the priority-date question, the travel-with-a-pending-I-485 question. Write the page that genuinely answers it: eligibility, realistic current timelines, what the agency actually asks for, what your firm does on those matters. Date it. Put an attorney's name on it.

1. Confirm it's indexed. Search Console → URL Inspection on that page. Crawled – currently not indexed means the work is invisible. Fix it before writing the next one.

1. Re-run your fan-out assumptions. Any keyword-to-prompt mapping your firm or your marketing vendor built before August 8 is working from a stale picture of how ChatGPT searches. The sub-queries changed materially.

The bigger picture

The industry spent two years optimizing to be findable. The behavior that changed on August 8 asks for something older and more familiar to attorneys: be the practice someone already thinks to call.

That is what a domain-scoped query encodes. It is a referral instinct, implemented in software — the engine going to the place it expects the answer to live rather than canvassing the open web and hoping. Firms have always built that position through consistency, specificity, and a reputation attached to a name. What's new is that the mechanism now reads your Google Business Profile and your indexed pages instead of taking a colleague's call.

One honest caveat: this is three weeks of data, and the people who collected it say plainly that they can show the timing and not the cause. Treat the direction as the reliable part. But the direction points somewhere small firms should welcome. When the engine goes looking for a named authority, a three-attorney practice that genuinely owns a question can be that authority. A content mill cannot.

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 whether AI assistants name your firm when clients ask — get your AI Visibility Score at CLIENTORY.ORG.