How to find the sources AI uses to recommend your competitors

Perplexity attaches numbered citations to every answer, which makes it the easiest engine to start with. ChatGPT shows source links on answers it produced with search on, and on some answers the searches it ran before answering; an answer given from memory shows neither. Gemini shows cited links when it grounds an answer in Google Search. Google AI Overviews and AI Mode carry their sources as link chips beside the summary and inside an expandable citation panel, so open the panel rather than trusting what is visible at first glance. Copilot lists the sources it shows beneath the answer.

Where no citations appear, the answer came from the model's memory, and you record it as such. A memory answer is still worth logging for whether you were named, but it names no pages, so it adds nothing to the source list. If an engine offers search on and off, use search on for this exercise and note which mode produced each answer.

Engines search more pages than they cite. A page the engine fetched and discarded did not shape the answer, and a list that mixes the two overstates your presence and understates the gap. Discoverable moved its own source lists and score from everything the engine searched to what the answer attributed on 8 August 2026; source lists shrank and the citation gap grew for definitional reasons, and readings either side of that date are different quantities. The dated history is on the methodology page.

By hand, the rule is the same. Record a page only if the answer attributed it. Where a surface shows the searches it ran, log them in a separate column; they tell you which questions the engine thought your prompt contained, which is useful later, but they are not citations.

Your own account colours the answer. Memory, custom instructions and past chats about your own company all change what comes back, so use a temporary chat or switch memory off, and stay signed out where the surface allows it. The reading to aim for is the public answer for the prompt at a given locale, without an account, which is what Discoverable's ChatGPT visibility tracker records for ChatGPT, brand-blind, with each reading labelled with the surface it came from. Repeat each prompt three times, because the citation set moves between runs, and a page cited in three of three runs matters more than a page cited once.

Set up a sheet with one row per cited page and these columns: prompt; engine; run (1 to 3); domain; page URL; content type (listicle, review platform, forum thread, news or trade press, directory, vendor page, documentation); names you (yes or no); names which rivals; and access. Access is whether you can get onto the page: open (you can create your own page or listing today), community (join and answer the threads where your category comes up), editorial (pitch the editor, leading with your data), paid (a listing costs money) or closed. Those are the classes Discoverable's own sources table uses, and they decide what you do next.

Fill it from ten prompts drawn from real buyer questions: category questions, comparisons that name your main rival, an alternatives question, use-case questions, and at most one branded prompt. Then sort by domain and count. The ranking you want is domains by the number of prompt-and-engine pairs that cited them. A review-site category page cited for six of your ten prompts on three engines is the most important page in your category, whoever it names.

Correct first. Where a cited page names you and gets a fact wrong, the fix is the correction procedure in how to fix wrong information AI gives about your brand, starting with the pages cited most often.

Open and community doors next. Review platforms want current, written, specific reviews from real customers, and do AI assistants read your reviews? covers what gets quoted. Forum threads want disclosed, useful answers over months; does Reddit improve AI visibility? sets out what works and what poisons the record.

Editorial doors last, because they are slowest. Listicles and comparison posts that omit you are pitched to their authors with a factual case and, where you have it, data they cannot get elsewhere. On paid plans Discoverable drafts those pitches as citation outreach drafts, which you edit and send from your own inbox; nothing is sent automatically.

Then keep the list current, because it moves. Reddit's share of ChatGPT Search citations fell from about 3.8 per cent to 0.5 per cent in the first half of August 2026 in one vendor's sample, and any single domain you depend on can move the same way. Re-run the ten prompts monthly by hand, or let the product do it: the citation gap counts the sources cited for rivals that never cite you, citation sources groups them by domain and content type, and the free AI visibility check returns a ranked citation gap with no login. The free check runs on Claude with live web search; paid plans also cover ChatGPT, Perplexity, Gemini, Copilot, Google AI Overviews and Google AI Mode, checking the core 25 prompts weekly on every surface and the full panel monthly.

If the list shows a rival named on pages where you are absent, why does AI recommend my competitors and not me? is the diagnosis. If you used to be on those pages and dropped out, why has ChatGPT stopped recommending my brand? works through the causes in order.