Why has ChatGPT stopped recommending my brand?

Open a temporary chat in ChatGPT, or switch memory off, and ask the prompt that used to name you five times, in five fresh chats. Count how many of the five name you. Do the same for four more prompts in the same family: the category question, an alternatives question, a comparison against the rival you lose to, and a use-case question. Twenty-five answers is enough to see a pattern. One answer never is.

Judge the result against your old baseline, if you have one. Named in four of five last month and one of five this month is a drop worth diagnosing. Named in one of one today is not evidence of anything. Discoverable checks weekly rather than daily for the same reason: a weekly reading of a fixed prompt set keeps single-answer randomness from reading as movement.

If the drop does not survive this step, the fix is a baseline. The free AI visibility check runs a set of buyer prompts about your brand through Claude with live web search, with no login; paid plans also cover ChatGPT, Perplexity, Gemini, Copilot, Google AI Overviews and Google AI Mode. Run it now, so that next month's question has an answer.

Your own ChatGPT account is not a neutral instrument. Memory, custom instructions, months of chats about your own company and your region all shape what you see. The reading that matters commercially is the public, logged-out answer for a prompt at a given locale. That is what Discoverable's ChatGPT visibility tracker records, brand-blind, with every report labelled with the surface it was read from; a temporary chat with memory off is the by-hand version of the same reading. If the drop appears only in your own account, it is a personalisation effect, and there is nothing on the web to fix.

Engines change retrieval without notice, and one change can move a whole category at once. The dated example is Reddit: between 18 July and 7 August 2026 it averaged 3.83 per cent of the links ChatGPT Search cited, and between 14 and 17 August 0.52 per cent, an 86 per cent fall in one vendor's sample. On 8 August the share of ChatGPT's background searches using a site: operator jumped from 0.37 per cent to 16.8 per cent. The numbers, and what they do not prove, are in ChatGPT has stopped citing Reddit.

The check is a comparison of cited domains. For each prompt in step 1, list the domains the answer cites now and compare them with what it cited a month ago. If the domains that used to carry your mentions have dropped out of the citations for everyone in your category, the engine moved, not you. If you kept no record, start one today.

A brand whose presence rested on two or three pages is exposed to those pages. A listicle gets refreshed and the new version drops you. A review profile's newest review turns two years old and the page stops being fetched. A comparison post is rewritten around this year's tools. Each removes the evidence the answer was built on, and the answer changes with it.

Open the pages the answer cites today and the pages it cited before, and read what each says about you and about the rival now named instead. The procedure for building that list is in how to find the sources AI uses to recommend your competitors. If the cited pages now describe a competitor in three sentences and you in none, you have been displaced rather than dropped; see why does AI recommend my competitors and not me?.

Two causes sit closer to home. The first is your own site. A robots.txt change that blocks OAI-SearchBot, the crawler behind ChatGPT search, a redesign that moved your pricing into JavaScript the crawler does not execute, or a pricing page moved to a new URL without a redirect will each take you out of the pages ChatGPT can read. The free AI crawler check reads your robots.txt against 12 AI crawlers in one pass; then load your key pages with JavaScript disabled.

The second is the model. Answers ChatGPT gives from memory, with search off, change when OpenAI ships a new model version, and nothing you publish reaches that memory until the next training run. Ask your prompts with search on and off. A drop that appears only with search off is a model change, and only the slow work of consistent third-party coverage answers it. A drop with search on points back to steps 3 and 4 or to your own site, and those are fixable.

On a weekly trend built from a fixed prompt set, a real drop is a step that holds: the score moves down and stays down across consecutive checks. A one-week dip that recovers is variation. On Discoverable, paid plans check the core 25 prompts weekly on every surface and the full panel monthly, and the free plan follows 25 prompts weekly on Claude with a monthly check across every surface. The Visibility trend plots the score per engine over time and the Response log keeps every captured answer. Alerts are on the roadmap; today you read the trend.

One rule prevents a false drop of our own making. When the definition of the score changes, each check is stamped with the methodology it was measured under, the chart marks the break, and no change figure is shown across it. The dated list of those changes is on the methodology page; never compare readings on opposite sides of a break. If ChatGPT never named you at all, the ladder is different: start with why doesn't ChatGPT mention my brand?.