Where do ChatGPT and Perplexity show their sources?
ChatGPT shows its sources as small citation chips inside the answer and, at the end of it, a Sources button that opens a side panel listing every page it used. Perplexity numbers its references in the text and lists the pages above the answer. Both cite only when they searched the web for that answer.
ChatGPT has searched the web and cited pages since OpenAI launched ChatGPT search on 31 October 2024. In some sessions since August 2026 the Sources button has moved into the three-dot menu on the answer, labelled View sources (Search Engine Roundtable, August 2026). If you see neither, ChatGPT answered from its training rather than the web, and there is nothing to record: ask it to search, or ask a question that needs current information.
Perplexity searches on every question and usually cites several pages, each with a number you can click. Its Sources list sits above the answer and opens in full with one click, so it is the quicker of the two to copy from.
The other assistants your buyers use show sources too, less consistently:
- Gemini: source links appear when the answer is grounded in Google Search, but not every answer shows them.
- Claude: with web search switched on, numbered citations link to each page it used.
- Google AI Overviews: the link icon beside each passage opens the pages it drew on.
Which questions should I ask?
Ask the questions a buyer types, not your company name. Ten is enough to start: three broad questions (“Who installs heat pumps in Exeter?”), three comparisons (“Vaillant or Mitsubishi heat pump for a 1930s semi?”), two problem questions (“Why is my heat pump so noisy at night?”) and two on price or place (“How much does a heat pump cost in Devon?”).
Write them as UK buyers write them, with the town, the trade term and the budget in. Run them in a private window, logged out where the assistant allows it, because history and saved memories shape answers. Note the date and the town you ran them from. Keep the list fixed between runs, or the results will not compare.
Add two questions that name competitors directly: “Is [competitor] any good for [service]?” and “[Competitor] or [you] for [service]?” The sources on those answers show you which pages the assistants hold about them, which is the comparison you came for. Our free Provley checklist has a ready set of ten questions to start from.
What should I record for each answer?
Record seven things: the question, the assistant, the date, every business named in order, every page cited with its domain, whose site each page is, and whether you were named, cited, both or neither. A spreadsheet with those seven columns is the whole tool.
| Column | What goes in it | Why it matters |
|---|---|---|
| Question | The exact wording you asked | Changing a word changes the answer |
| Assistant and date | ChatGPT, Perplexity, Gemini or Claude, and the day | Each assistant draws on different sites |
| Businesses named | Every name, in the order given | First place is what a buyer reads |
| Pages cited | Each address, with its domain | The domain tells you the type of source |
| Whose site | Yours, a competitor’s, a directory, press, a forum | Third-party sites are winnable |
| Result | Named, cited, both or neither | Each needs a different fix |
| Wrong facts | An old address, a closed service, a wrong price | Errors need correcting at the source |
Named and cited are different results. Cited but not named means your page supplied a fact while the answer recommended someone else; the fix is to put your name, service and area into the passage it quotes. Named but not cited means a directory or review site carried you; the fix is to strengthen your own page so the assistant can confirm the details directly. Why ChatGPT and Perplexity recommend your competitors but not you works through each pattern.
How do I compare my sources with my competitors’?
Sort every cited domain into a type, count how many answers cite it, and read the lists side by side: their own pages, the third-party sites that mention them, and the questions they win. The comparison is domain by domain, not score by score.
The types worth separating for a UK business are these:
- Their own site, and which pages on it
- Directories and review sites: Google Business Profile, Checkatrade, Trustpilot, Feefo
- Trade bodies and registers: Gas Safe, NICEIC, the Law Society, the FCA register
- Press and trade magazines
- Forums and social: Reddit, Mumsnet, LinkedIn
- Reference sites: Wikipedia and government pages
Forums matter more than most owners expect. In Semrush’s 2025 study of roughly 150,000 citations across ChatGPT, Perplexity, Google AI Mode and AI Overviews, Reddit was the most cited domain at about 40% of citations, with Wikipedia second at about 26% (Semrush, Most cited domains in AI answers, 2025). A thread where a buyer in your town asked for a recommendation can outweigh your own service page.
With the domains sorted, ask three questions of the competitor’s list:
- Which of their own pages get cited, and what do those pages do? Open each one. Does it answer in the first lines, carry a dated figure with a named source, show a date, and end with an FAQ? Those are the traits the assistants reward, and they are copyable.
- Which sites cite them but do not mention you? A listing, a review profile or a trade body page they hold and you do not is a winnable mention: you can hold it too.
- Which questions do they win that you do not? Start your fixes on the pages behind those questions.
Being cited pays twice. Seer Interactive’s September 2025 update found that brands cited inside a Google AI Overview had about 35% higher organic clickthrough across the third quarter of 2025 than brands that were not cited, so a page the assistants quote also earns more of the clicks that remain.
How often should I repeat the check?
Weekly, with the same questions, on the same day, and more than one run per question. Answers change between runs, assistants and towns, so a single check is a snapshot, not a measure. The number to track is your Citation Share: the share of answers that name your business. Expect it to move over weeks rather than days, because the assistants have to re-read your pages and directories take time to update.
What does Provley do that a spreadsheet does not?
Provley asks ChatGPT, Perplexity, Gemini and Claude your tracked questions every week from its own accounts, keeps every answer with every page it cited, and turns the record into two views you would otherwise build by hand.
- Sources: every domain cited across your answers in the last 30 or 90 days, ranked by the share of answers citing it; your own pages cited and your share of citations; citations by type of site; and winnable mentions, the sites the answers trust that do not name you yet.
- Competitors: every site named for your questions, with its visibility (the share of answers naming it), its share of voice and its average position; the pages the assistants cite most for it; the questions it wins; and its most cited page set against yours on word count, whether it answers first, how many figures it carries, whether it shows a date and whether it has an FAQ.
Provley’s weekly AI check tracks ten questions on the plan, £149 a month with 6,000 credits, and the Monday email carries your Citation Share and the next page to fix. Copilot and Google’s AI Overviews cannot be checked automatically, so Provley records them by hand. If you want the measure before the plan, the £29.99 audit scores ten of your pages for AI and for Google.
Frequently asked questions
Do I need a paid account to see ChatGPT’s sources?
No. ChatGPT search, with its citations and Sources panel, has been available without an account since February 2025. Logged-out answers can differ from logged-in ones, so record which you used and keep to it.
Is being cited the same as being recommended?
No. A citation means the assistant used a page as evidence. A recommendation means it named a business. Your page can be cited while a competitor is recommended, and a directory can get you recommended without your site being cited. Record both, because the fixes differ.
Will the sources be the same tomorrow?
No. Assistants retrieve pages afresh for each question, and the set changes with wording, place and day. That is why the check is weekly and why Provley keeps every answer rather than a score: the answer is the evidence, and the trend is the measure.
Can I see the sources an assistant used for a question about a competitor?
Yes. Ask the question as a buyer would and open the same panel. Nothing about a competitor is hidden. What you learn is which of their pages, and which third-party sites, the assistant relies on for them, and that list is your to-do list.