DeepSmith

Nov 26 · AEO & AI Visibility

20 min read

How to Find the Sources AI Engines Cite Most in Your Industry

Avinash Saurabh
Avinash Saurabh · CO-Founder & CEO
Abstract monochrome cover showing small document cards on the left connected by thin lines to a few larger circular source nodes on the right, around the cover line Find the Sources AI Cites.

You've probably seen a few familiar websites show up in an AI answer about your category, and it's hard to tell whether those sites keep coming back or whether you just happened to see them once. This guide is for marketing leads who want to replace that hunch with a source list built from a repeatable set of questions. If you want a quick tool shortlist first, the best AI citation tracking tools roundup covers that side, and this guide covers the method. By the end you'll have two separate rankings, one of cited domains and one of cited pages, along with the prompts and saved answers behind every count.

You'll need a spreadsheet, access to the AI search tools you want to check, and a clear idea of the category you care about. Working out the sources answer engines cite in my industry is a form of industry citation analysis, and the result describes what the engines showed in your sample. It isn't a full list of what an engine reads or was trained on, and it doesn't tell you why a source was picked.

Step 1: Define the category and the collection rules

Before you open any AI tool, write a short scope statement. It should name the category, who is asking, the geography and language, the engines, the mode on each engine, the collection dates, and what will count as a source. Here is an illustrative one: "English-language answers about expense management software for UK finance teams at companies with 50 to 250 employees, collected through ChatGPT Search, Perplexity, and Gemini Apps on three separate dates." It only shows the shape and isn't a finding about that market.

Treat each engine as its own surface: ChatGPT Search, Perplexity, Gemini Apps, Google AI Mode, and Google AI Overviews. Gemini Apps is not Google AI Mode, and AI Overviews is not AI Mode either. Google's documentation says AI Overviews and AI Mode may use different models and techniques, and may fan a question out into related subtopics and sources, according to its AI features documentation. If Copilot or Claude matters to your audience, there are tools that track which sources Bing Copilot cites and tools that track which sources Claude cites, which you can add later.

For a first pass, pick three surfaces your audience really uses, and don't say they stand in for all AI usage. A workable starting design is 15 distinct prompts, three engines, and one run of each prompt on each engine across three separate dates in the same week. That's 15 × 3 × 3, or 135 planned answers, which is my suggestion for a manageable pilot and not a statistical minimum.

You're done when another person could read your scope and collect the same data without choosing a different audience, region, engine, or mode.

The usual mistake is mixing a consumer question, a local service search, and an enterprise buying question into one leaderboard and calling it "our industry." Another is combining ordinary search, deep research modes, and long follow-up chats as if they were the same test.

Step 2: Build a balanced seed prompt set

A seed prompt is a fixed question you use to sample answers about your category. Pick questions that reflect what people in your market really need to know, not questions designed to make a favorite site show up.

Use five intent families with three prompts each:

  • Understand the category: "What is expense management software?"
  • Solve a practical problem: "How do finance teams handle missing expense receipts?"
  • Evaluate options: "What should a finance team compare when evaluating expense management software?"
  • Investigate requirements: "What approval controls matter when choosing expense management software?"
  • Research costs and implementation: "What questions should buyers ask during an expense management software demo?"

Those are illustrative examples, so swap in questions that fit your own category. Sales calls and support tickets are good raw material.

Keep a small prompt register with an ID, the exact wording, the intent family, and a version date. Freeze the wording for the first round, and if you change a question later, keep the old one and give the new one a new version ID. Don't add "use Wikipedia" or a preferred publisher's name to a prompt, because that turns the test into one where you steer the sources instead of watching which ones the engine picks.

DeepSmith's Discover Prompts can suggest questions from your product and persona context, which makes a decent first draft. Edit it for balance and take out anything that leans toward your own brand. Treat the suggestions as a starting point, not a measure of search demand.

You're done when you have 15 clearly different questions across all five families and the exact wording is saved.

People often write 15 versions of "best software," which leans toward commercial recommendations and can miss documentation, standards bodies, and research publishers.

Pro tip: If most of your prompts contain "best" or "alternative," swap some for definition and implementation questions first.

Step 3: Run the prompts under consistent conditions

Put each saved prompt into a fresh conversation or a separate search, and repeat the same wording on each planned date. Hold these steady as far as the interface lets you: language and location, logged in or out, search or research mode, model choice, connected apps and files, memory settings, and time of day.

A fresh conversation cuts carryover but doesn't prove personalization is gone. OpenAI documents that Memory can influence how ChatGPT writes search queries, so write your settings down and don't call the test neutral.

Here is what to do on each surface:

  • ChatGPT Search: open the tools menu and pick Search, or type a slash and choose Search. Search can turn on by itself, as the ChatGPT Search help page explains, but choosing it on purpose makes the condition clearer, so note whether it ran.
  • Perplexity: submit the question in one chosen mode and look at the numbered citations. Don't mix standard search with advanced research results unless you label them.
  • Gemini Apps: open Sources where it appears, inline or below the response, and keep supporting links apart from related ones.
  • Google AI Mode: submit the question and look at the web links tied to the response, separate from Gemini Apps and from ordinary results.
  • Google AI Overviews: run the query in Google and note whether an Overview appears. It isn't guaranteed to, and ordinary results shouldn't be swapped in when it doesn't.

Save the full answer, the collection time, and an answer ID. Screenshots help you remember where a source sat, but the destination URL is what goes into your data.

Give every answer one status: completed with eligible sources, completed with no visible sources, completed with only related or unclear links, failed or blocked by a limit, capture incomplete, or AI Overview did not appear. A completed answer with no visible sources is a real observation of zero. A failed request is missing data, and it isn't a zero.

This is where doing it by hand gets tedious. DeepSmith can run tracked prompts on a schedule and keep the full answers for the engines your plan includes, so you aren't repeating the collection yourself, and you can open the saved answers to check the evidence behind a count. Not every plan covers every engine. If content teams are weighing this trade-off, there are citation tracking tools for content teams to compare.

You're done when every planned prompt, engine, and date combination has an answer record or a note explaining why it's missing.

The classic mistake is rerunning a question until a familiar source appears, keeping that one, and dropping the rest. That is selective sampling, so keep every planned observation and document any replacement.

Step 4: Extract the sources actually attached to each answer

Open the citation controls and copy the destination page, not just the publisher's name. In ChatGPT the Sources control under an answer can show both cited sources and other relevant links, so don't assume every item has the same standing. Perplexity documents numbered citations that link to original sources, so read those instead of asking the model to write a bibliography afterward. In Gemini, Google's help pages call them "sources and related content," so if you can't tell which a link is, label it unclear.

Gemini's Double-check feature is a different thing: Google says the page it links is not necessarily what Gemini used to write the answer, so keep those links out of your main tally.

Sort every link into one of three classes:

  1. Attributed: attached to a claim in the answer, or clearly shown as supporting it.
  2. Related or verification: offered for further reading or checking, without a clear tie to how the answer was built.
  3. Unclear: the interface doesn't make the relationship clear.

Use the attributed links for your main ranking and keep the other two in a side view. Leave out ordinary search results outside the AI response, ads, navigation links, and private files. Asking an engine to name the most authoritative sites in your industry doesn't give you an observed record either. If you'd like more detail on what counts as a citation, a separate guide explains what AI citation tracking records.

For each source, capture the answer ID, prompt ID, engine and mode, timestamp, raw destination, page title, link class, citation position, normalized page, registered domain, source type, and a notes field for redirects, errors, or odd destinations.

You're done when every included source traces back to a saved answer and a visible attribution, and the related and unclear links are labeled and not hidden.

Common mistake: A source panel isn't automatically a list of every page the engine read, and a verification link isn't proof of the page that produced the answer. A brand named in the answer text isn't necessarily cited either, because this guide counts source references and not mentions.

Step 5: Normalize the destinations and remove duplicate appearances

Keep the raw URL exactly as you saw it and add a second column with a cleaned version, so the original evidence survives the tidying. Clean conservatively:

  1. Follow redirects where you can and note the final destination.
  2. Remove fragments that only point to a spot within the same document, and tracking-only parameters that don't change the content.
  3. Keep parameters that point to a different article, product, language, or version.
  4. Lowercase the host, but not the whole path, since some sites treat path case as meaningful.
  5. Merge versions only when you can show they land on the same content. Don't assume every trailing-slash variant, mobile page, translation, or syndicated copy is identical.
  6. Work out the registered domain with public-suffix-aware logic or by hand, because taking the last two labels of a hostname gets country-code domains wrong.

Decide your domain grouping rule once and use it everywhere. For a small pilot, doing this by hand is fine.

Now build two tables. Domain appearances has one row for each unique answer ID plus registered domain, and page appearances has one row for each unique answer ID plus normalized page. A page cited three times in one answer counts once. Two pages from one domain in one answer give two page appearances but one domain appearance. A domain in three separate answers gives three domain appearances. Deduplicate inside each answer only, because removing a repeated domain across the whole workbook wipes out the recurrence you're trying to measure. The difference between the two granularities is covered in the guide on domain-level and URL-level citation tracking.

You're done when the domain table has no repeated answer ID and domain pairs and the page table has no repeated answer ID and page pairs.

People either count every repeated citation marker and inflate the numbers, or remove a domain across the whole dataset and lose the recurrence.

Step 6: Tally domains and pages with explicit denominators

Make one ranking for domains and another for pages, and keep the main calculation simple: how many completed answers contained this source?

  • Appearances: the number of distinct completed answers with at least one eligible reference to the domain or page.
  • Prevalence: appearances divided by completed answers in the sample, times 100.
  • Prompt breadth: the number of distinct prompt IDs the source appears for.
  • Date breadth: the number of distinct collection dates it appears on.

Work these out for each engine first. A combined ranking is optional, and if you make one, say how the observations were combined. Completed answers with zero sources belong in the denominator. Failed or incomplete captures stay out of it, but show how many there were. For AI Overviews, report attempted searches, Overviews that appeared, and Overviews captured on separate lines.

Here is a worked example with made-up data. Take six completed answers from one engine in one mode, where A4 has no eligible sources.

AnswerAttributed pages observed
A1Directory A, page D1; Directory A, page D2; Trade journal B, page J1
A2Directory A, page D1; Vendor C, page V1
A3Directory A, page D1; Trade journal B, page J2
A4No eligible attributed sources visible
A5Trade journal B, page J1; Vendor C, page V1
A6Directory A, page D2; Trade journal B, page J1

The domain ranking comes out like this:

Domain labelAnswers containing itPrevalence across six answers
Directory A466.7%
Trade journal B466.7%
Vendor C233.3%

Directory A counts once in A1 even though two of its pages were cited. The denominator stays at six because six answers were completed, even though only five had sources. If someone divides Directory A's four appearances by those five, they get 80%, a different, conditional number that shouldn't replace 66.7% without saying so. On the page side, D1 and J1 each appear in three answers (50%), D2 and V1 in two (33.3%), and J2 in one (16.7%). There are eleven answer and page pairs but only ten answer and domain pairs, which is why a domain list can't stand in for a page list.

In a spreadsheet, keep an answers tab with one row for every planned observation, including statuses and zero-source answers, and a pivot table with the domain in rows, the engine in columns, and a count of answer IDs as the value. Pull denominators from the answers tab, not the source table, because the source table has no rows for zero-source answers and would push every percentage up.

When you'd rather not rebuild this by hand, DeepSmith has cited-source views for domains and for pages. Its get_cited_sources capability returns the most-cited domains or URLs across your tracked prompts, tagged as your own, a competitor's, or neither, with a range of 7, 30, or 90 days (30 by default) and up to 50 results. One boundary matters here: that view aggregates across every tracked prompt, and nothing documented gives it a filter for one seed set or one engine. If your workspace also holds unrelated prompts, build the scoped tally from the underlying answer records, which DeepSmith returns with full text and citations through CSV and API export, instead of presenting a workspace-wide leaderboard as category-specific. A product's displayed total may not use this guide's within-answer deduplication, so keep its label or recalculate from answer-level records. If you're still choosing tools, a roundup of source-finding tools compares them.

You're done when every percentage has a named denominator, every ranked source traces back to its answers, and the domain and page results are kept apart.

A common slip is comparing raw citation totals from engines with different numbers of answers. A preprint on measuring AI visibility makes the same point: raw citation counts scale with answer volume and with how many citations a platform emits, while prevalence measures the fraction of responses containing a domain, however often it repeats within one.

Step 7: Inspect the recurring domains and the exact pages

Use the domain ranking to spot recurring publishers, then the page ranking to see which documents are behind that recurrence. Add a short source-type label to each, such as official documentation, trade publication, review directory, standards body, or community platform. This is your own classification and not a credibility score.

For the leading pages and any odd entries, write down what the page really is, which prompt families brought it in, whether it loads or redirects, whether the cited material looks relevant, and whether it's a primary source, a review, or a discussion. If an answer visibly cited a broken or questionable page, it stays in the dataset with a flag, because whether something gets cited is a separate question from whether it's reliable.

Then give each source a working label:

  • Recurring across questions and dates: appears for at least two prompts and on at least two dates.
  • Intent-specific: appears repeatedly inside one narrow prompt family.
  • Single observation: appears in only one answer.
  • Attribution unclear: the visible link doesn't support a confident call.

The two-prompt, two-date rule is a suggestion for staying organized and doesn't prove significance or that a source will keep appearing. Look at the engines side by side before relying on a pooled total, since a source can recur on Perplexity and be missing from your Gemini answers. The GetMentions citation-volatility study, vendor-published research from one platform's data and a seven-day window in June 2026, ran 2,398 queries across four surfaces and found that about 84% of the roughly 72 domains cited per query were seen on exactly one engine. The authors call this directional, so use it as a reason to keep engines apart and not as a number that holds in your industry.

You're done when each leading domain has an identifiable page and prompt context behind it, and a colleague can tell broad recurrence from a source that matters for one question only.

Calling a frequently cited site "trusted by AI" is a mistake, as if you'd measured an internal trust score. Say "frequently cited in our sample," and stop at understanding the landscape.

Step 8: Save a dated source list and repeat the same test

Make two deliverables, a domain list and a page list, each with appearances, the denominator, prevalence, a per-engine breakdown, distinct prompts and dates, source type, and notes. Put the scope statement, collection dates, prompt set version, failed request count, and eligibility rule with both, since a short methods header is far more useful than an unexplained leaderboard. That header is what makes your industry citation analysis something a colleague can check.

A reasonable rhythm is to refresh the frozen core set monthly and review the seed set every quarter or when your category changes. That's my recommendation, not a proven best interval. If the ranks move a lot between rounds, collect more matched observations before making strong claims. Keep old snapshots, and mark the break whenever your prompts, modes, or engines change.

For context, Pew Research Center's 2025 analysis of Google AI summaries found Wikipedia, YouTube, and Reddit were the most cited sources, together 15% of the sources listed in AI summaries. That was a historical US Google study, so it shows that a few recognizable domains can stand out in a big sample without being your industry's source list.

You're done when you have a dated, reproducible source list, an owner, and a date for the next collection, and another analyst could rebuild the ranking from the answers you kept.

The usual mistake is passing the list around without its prompt set or collection window and then treating it as a permanent map of the industry.

What to do next

Keep the prompt set and collection rules in the same file as your source list. Repeating the same test gives your team a clearer view of which domains and pages recur, without confusing one answer with an industry-wide pattern. Marketers comparing platforms can start with the citation tracking tools for marketers, and B2B teams have a comparison of citation tracking tools for B2B marketers. The wider AI visibility tools roundup covers the category. Even a light version of AI chatbot citation sources analysis becomes more useful after two or three saved rounds.

Want to make the collection repeatable? Start a DeepSmith 7-day free trial to set up prompts, capture answers, and inspect the sources behind them.

Frequently asked questions

How do I find which sources AI engines cite most in my industry?

Run a fixed set of category questions on the engines you care about, repeat them on separate dates, and record the sources attached to each answer. Clean the destinations, count each domain once per answer, and build a separate page tally. Report the prompts, collection window, and denominator along with the rankings.

How many prompts and repeated runs do I need?

A practical first pass is 15 prompts across five intent families, collected on three engines on three dates, which gives 135 planned answers. That's a manageable pilot and not a statistically proven minimum. Expand it if important intents are missing or the leading sources keep changing between observations.

Should I rank domains or individual pages?

Rank both. The domain list shows which publishers or platforms keep recurring, and the page list shows which exact documents appear. Several pages from one domain can show up in the same answer, so the two counts won't always match.

Does every source panel or verification link count as a citation?

No. Some interfaces mix cited sources with related links, and Gemini's Double-check links, for example, aren't necessarily pages used to write the original answer. Use clearly attributed sources for the main tally and keep related or unclear links separate.