Your own site is a minority of what AI answers cite in most categories. Map the third-party sites behind your buyer prompts before spending another week polishing pages that never enter the answer.

The concentration can be awkward. In one count of 680M citations, Wikipedia accounted for 7.8% of ChatGPT citations, while Reddit accounted for 6.6% of Perplexity citations. Those figures are not your category map. They are a reason to build one.

I used to tell clients to fix their own pages first. I was wrong about the order. Fix your pages, yes. But first identify which third-party sources already shape the answers. That list tells you where to earn a mention.

By the end of this recipe, you will have:

  • A ranked table of cited domains, grouped by type and split by engine where available.
  • The cited pages behind the top five third-party domains.
  • Five concrete jobs, with owners and deadlines.

Before starting, have Claude with groas connected, 8–12 buyer prompts already tracked in groas, and 30 minutes. No tracked prompts yet? Run the afternoon workflow for testing buyer prompts, then return.

Use buyer questions, not searches for your brand.

1. Select 8–12 prompts that cover buying decisions

Open your tracked prompts in groas. Select a set that covers comparisons, problems, and reviews. I would use 10 prompts with this split:

  • 4 comparison prompts: best crm for small agency, top project software for contractors.
  • 3 problem prompts: how to reduce no-shows for cleaning business and equivalent questions in your category.
  • 3 review-intent prompts: g2 vs capterra for hr software reviews and category-specific requests for Reddit opinions.

Use examples from your own category, not this list verbatim. Keep your brand name out of the set.

Citation mix changes with intent. For explicit review queries, 49% of AI Overviews cited a review platform, versus 17.1% for best/top queries. A set made entirely of comparison prompts answers a narrower question than a set covering the buying decision.

Expected result: one saved set of 8–12 non-brand buyer prompts, with all three intent groups represented.

Mistake to avoid: tracking what you wish buyers asked. A report full of searches for your brand can look reassuring while telling you little about where buyers encounter alternatives. Build the category map, not the cleanest-looking report.

2. Pull the citation records into Claude

In Claude with groas connected, paste the following. Replace the bracketed fields before sending.

Using groas citation data for my tracked prompt set [SET NAME] over [DATE RANGE]:

1. List every available citation record behind those answers.
2. Include the prompt, cited URL, domain, and engine where reported.
3. Count citation records per domain. Do not deduplicate across prompts.
4. Separately count the distinct prompts where each domain appears.
5. Give each domain one primary type: owned, competitor, forum, review,
   video, editorial, media, blog, or other.
6. Keep my own domain, [YOUR DOMAIN], labeled owned.

Return a table:
domain | type | citation count | distinct prompts cited in |
engine | example URLs (2 max) | prompts where cited

Flag missing fields rather than guessing them.

Get the count table before asking for interpretation. A summary can tell you that Reddit matters. It cannot tell you whether Reddit appeared behind one prompt or nine.

Keep two measures separate:

  • Citation count: how often the domain appears in the citation records.
  • Prompt coverage: how many distinct prompts include that domain.

Both belong in the table. Repeated citations behind one question are not the same as appearing across the whole prompt set.

Expected result: a domain table you can trace back to prompts and example URLs. Your own domain may sit well below the leaders. Do not force a particular number of domains or a preferred ranking.

Mistake to avoid: asking Claude for “the main insights” first. Summaries compress the very repetition this exercise needs to preserve. Count first. Interpret second.

3. Separate engine rankings and label core sources

Continue in the same Claude chat so the records and counting rules stay together. Paste:

Using the citation records above:

1. Split results by engine wherever groas reports the engine.
2. Within each engine, group domains by type, then sort each group
   by citation count descending.
3. Show distinct prompt coverage as both a fraction and a percentage.
4. Use the number of tracked prompts with available answers for that
   engine as the denominator. Show that denominator.
5. Label a domain core source when it appears in 30% or more of
   those prompts.
6. Put records without an engine label in a separate group.

Do not invent rankings for engines with no available data.

Treat core source as a working label, not a badge of authority. Here, it means a domain appears across enough of your prompt set to deserve attention.

Keep the engine split visible. A blended ranking can hide where a source actually matters. In the 1,274-run AI Overviews sample, YouTube appeared in 25.2% of runs and Reddit in 18.2%. Your own set still has to establish its own order.

Expected result: a separate ranking for each engine represented in your data, with consistent domain labels and explicit denominators.

Mistake to avoid: treating a competitor blog cited twice as equivalent to a forum cited across nine prompts. Check coverage before deciding where to work.

4. Rank 15 third-party domains, then inspect five

Paste this follow-up:

From the same records, rank the top 15 third-party domains by total
citation count. Exclude domains labeled owned.

Show:
domain | type | citation count | distinct prompt coverage |
engine breakdown | core source status within each engine

For the top 5 domains:
- List one cited page URL per domain.
- Identify the prompts and engines that cited that page.
- Show relevant passages only where the available data contains them.
- Distinguish answer text from cited-page text.
- Do not describe a page passage as a quotation used by the AI answer
  unless the available answer text supports that description.

Flag any page I need to open manually.

The overall top 15 is a shortlist. The engine breakdown keeps it from becoming a misleading average.

Open the five cited pages. Read them. Look for the specific block that could earn a mention: a Reddit comment with a price, a comparison-table row, a transcript passage, or a documentation paragraph that states a limit plainly.

A citation alone does not show which sentence an answer used. If the answer text is available, compare it with the page. Otherwise, label the passage as a candidate, not a proven extraction.

Review platforms deserve the same scrutiny. In software answers tracked from September 28 to October 4, 2026, G2 earned 0.73% of ChatGPT citations and Capterra 0.32%. Review sites plus analysts accounted for 1.57%, against 2.64% for docs/help and 2.01% for vendor sites. Do not let a familiar review-site logo choose the work for you. Read your ranking.

Clay models of Reddit, YouTube, and review-site icons stacked by size on a desk map

Expected result: five domains, five cited pages, and a relevant passage or a clearly marked manual check for each.

Mistake to avoid: stopping at a one-off checker. A tool that checks AI Overviews, ChatGPT, and Perplexity live for one query can demonstrate a citation. The deliverable here is different: a tracked prompt set, domain rankings, type labels, and a repeatable comparison. Keep the groas table for the decision.

5. Turn the top five into jobs due within 14 days

Give Claude the owners and dates rather than letting it invent a staffing plan. Paste:

Turn the top 5 third-party domains into five actionable jobs.

Available owners: [NAMES OR ROLES]
Deadline: [DATE WITHIN 14 DAYS]

Return:
domain | cited page or thread | what to publish or fix
(under 20 words) | owner | due date

Prefer jobs that create a specific, citable block:
price, limit, comparison row, steps, or timestamp.

Use only facts already available. Flag missing facts.
Distinguish edits we control from contributions or corrections
that require someone else's approval.

If a domain has no realistic job within 14 days, flag it and suggest
an actionable replacement from the top 15. Do not pad the list.

Cap the list at five jobs. Forty rows are not more strategic. They are easier to leave untouched.

Choose the job according to the source:

  • Forum: answer the exact question in the cited thread, where a useful contribution is possible. Include numbers and a drawback. Do not pitch. A detailed answer gives the reader something to use; ten brand mentions give them a reason to ignore you.
  • Review platform: fix profile completeness first, then recency. Fill required fields, add current customer quotes that state a use case, and reply to recent negatives. Do not make chasing half a star the entire assignment.
  • Video: publish or sponsor a focused video that says the useful sentence out loud. Use chapters so the relevant section is easy to find. Write down the fact the video needs to communicate before writing the brief.
  • Editorial or media: pitch a specific correction to an existing relevant article, not a vague feature idea. Identify the outdated facts and supply a table row the writer can use. The writer still decides whether to use it.
  • Competitor documentation or blog: copy the useful structure, not the words. If the cited page clearly presents limits, pricing tiers, and steps, check whether your page states those same facts plainly. This job improves your owned page; it does not earn an edit on the competitor’s site.

Expected result: five rows with a concrete deliverable, a real owner, and a deadline. Each row should say what changes, not merely what someone researches.

Mistake to avoid: assigning a job that depends on access you do not have. A national publisher requiring a long PR engagement is not a 14-day task. A profile correction or a useful contribution to an existing discussion may be.

For the broader allocation question, use the method for finding which sources AI cites in your industry. This recipe supplies the category-specific work queue. Assign it before expanding it.

6. Re-run after 30 days without changing the test

Return to the same Claude chat after 30 days. Paste:

Re-pull groas citation data for the same tracked prompt set using
a new date range of the same length.

Keep the same prompts, domain-type labels, and counting rules.
Compare engines separately wherever engine data is available.

Show:
- New domains in the top 15.
- Domains that fell out of the top 15.
- Changes in distinct prompt coverage for each core source.
- Prompts where my own site gained or lost a citation.
- Changes in available answer data that affect the comparison.

Where answer or cited-page text is available, flag whether the
specific block we published or corrected now appears.
Do not attribute ranking changes to our work without evidence.

Hold the prompt set steady for 90 days. Change the targets, not the test.

Watch source coverage and the actual content of the answers. If Reddit falls from 7/10 prompts to 4/10, that is movement, not proof of progress. Check what replaced it and whether the contribution you made appears. A disappearing source is not automatically a win.

Also flag prompts where your own site remains absent despite a relevant fix. The sources-cited gap view describes this useful distinction: which sources appear where yours does not. Use those prompts to choose the next investigation.

Expected result: a short movers table, plus a clear distinction between observed changes and confirmed appearances of your work.

Mistake to avoid: rewriting the prompt set because the numbers look flat. That gives you a different map, not evidence that the original work succeeded or failed.

Verify the map, then ship the highest-priority job

You are done when you can answer these questions without reopening Claude:

  1. Which five third-party domains appear most often behind the selected buyer prompts?
  2. What type is each domain, and on which engines does it matter?
  3. Which cited page and specific passage make each one actionable?

Save the table and chat. If a count looks wrong, trace it to the underlying citation records. Then spot-check a prompt in ChatGPT or an AI Overview against what groas logged. Counts can drift between runs, so do not expect a fresh answer to reproduce every recorded citation exactly.

A map you can assign is done. A dashboard you admire is not.

Once the workflow works, change one thing first: pick the actionable core source with the widest prompt coverage and ship its job this week. That may mean posting the detailed forum answer or completing the review profile. I would skip pitching a new editorial feature until that job ships.

Re-pull the same prompt set in 30 days. Check that domain’s coverage, then inspect whether your contribution or corrected information appears. Do not call a quieter domain a success without reading the answer. Use what you find to choose source two. That is the loop.