When ChatGPT, Gemini or Google AI Mode answers a question in your category, it usually leans on a short list of third-party pages: a comparison post, a "best tools" roundup, a forum thread, a review. If your brand is missing from those pages, or described wrongly on them, the AI answer tends to miss or misdescribe you too. You can publish more on your own site, but the faster lever is often the pages the AI already trusts.
AI citation outreach is the work of finding those cited pages, deciding which are worth pursuing, and persuading their owners to include or correct your brand. It is classic link-building outreach with a different target list and a different success measure: not a link for PageRank, but a fair, accurate mention on a page an AI engine reads.
This guide sets out a working process using Semrush's Prompt Tracking Sources report to find cited pages and Claude Code to triage hundreds of URLs and draft first-pass research and pitches. It covers the prompt set, the export, the Claude setup, a scoring model, pitch templates for each type of opportunity, the line you should not cross, and how to tell whether any of it moved the answers.
Why a cited page that never names you is an outreach target
Semrush and Kevin Indig's ghost citations study (June 2026) looked at 3,981 domain appearances across 115 prompts, 14 countries and four AI search engines. In 61.7% of appearances the AI linked to a brand's page as a source but never named the brand in the answer. Only 13.2% were both cited and named, and 25.1% were named without a link.
The engines behave very differently. ChatGPT cited sources 87% of the time but named brands in only 20.7% of answers; Gemini named brands in 83.7% of appearances but cited sources only 21.4% of the time. And intent matters: informational queries had an 89.3% citation rate but an 18% mention rate, while comparative queries had a 43.3% mention rate.
“There's almost no overlap between which brands ChatGPT cites and which ones Gemini names for the same prompt.”
— Kevin Indig, quoted in Semrush, Why 62% of AI citations don't lead to brand mentions [Study]
Two conclusions shape the workflow. First, the pages that get cited are often not the brands that get named; a cited roundup that lists your competitors and not you is exactly where a named mention can come from. Second, comparative and evaluation prompts are where naming happens most, so that is where outreach pays off first. If you want the underlying reasons AI picks other sites over yours, read our explainer on why AI cites other sites instead of yours before you start.
Links still play a part too. A separate Semrush study of 1,000 domains (October 2025) found a Pearson correlation of 0.65 between domain authority and AI mentions, and nofollow links correlated almost identically to follow links (0.340 versus 0.334). Correlation is not causation, but it supports treating a quality mention on a trusted page as valuable even when the link is nofollow.
Building the prompt set that decides which pages you chase
Your target list can only be as good as the prompts that generate it. Semrush's own guide to AI citation outreach advises tracking both awareness and evaluation-stage prompts and avoiding prompts that include your brand name. Branded prompts tell you how AI describes you; they do not show you the pages that decide whether you appear at all.
Prompt Tracking limits are tight, so choose deliberately. According to the Semrush knowledge base, the AI Visibility Toolkit alone allows 25 prompts per account, Semrush One Starter 50, Pro+ 100 and Advanced 200. A balanced 25-prompt set for a B2B software brand might look like this:
- 8 category prompts: "best invoicing software for freelancers", "top tools for recurring billing".
- 7 problem prompts: "how to stop clients paying invoices late", "how to automate payment reminders".
- 6 comparison and alternatives prompts: "[competitor] alternatives", "[competitor A] vs [competitor B]".
- 4 qualifier prompts: "invoicing software with multi-currency support under $20 a month".
Pull the wording from sales calls, support tickets and forums rather than keyword tools alone; Semrush's Sources report guide recommends the same. If you sell to a buying committee, map prompts to each role, as covered in our guide to B2B answer engine optimization.
Pulling cited pages out of Semrush's Sources report
Prompt Tracking supports ChatGPT Search, Google AI Mode and Gemini, and the Sources tab lists cited domains and pages with per-page mention details (Semrush KB). The steps:
- Create a Prompt Tracking campaign for your domain, choose the AI platform and location, and add your prompts. Semrush notes data can take up to 24 hours to populate.
- Open the Sources tab and switch platforms with the dropdown at the top to compare the same prompts across engines.
- Use the Sources by Category widget to see which source types each platform relies on, such as retail or knowledge bases.
- Switch to the Pages view to see the specific URLs cited most often.
- Sort for pages with a high Cited for prompts count and a low Mention rate. These are your first candidates.
- Use Get mentions beside a page to see which prompts cite it and whether your brand appeared.
- Export to CSV (one file per platform is easiest to work with) and save it in a project folder.
| Sources report metric | What it tells you | How to use it in outreach |
|---|---|---|
| Cited for prompts | How many tracked prompts cite this page | Higher means one win affects more answers |
| Your mentions | Answers citing this page that mention your brand | Zero on a high-citation page is a priority gap |
| Mention rate | Share of associated answers that mention your brand | Measures the answers, not the page: you can be on the page and still absent from the answer |
| Diff | Change since the last update | Spot pages that recently started or stopped being cited |
| Domains view | Prompts where each domain is cited | Benchmark against competitors and find publishers worth a relationship, not just a page |
Note the caveat Semrush's own Sources guide makes: mention rate describes the AI answers. A low rate might mean you are missing from the page, or that you are on it but buried. Only reading the page tells you which, and that is where Claude helps.
Setting up Claude Code to triage hundreds of cited URLs
A single export can be large. In Semrush's worked example, one run produced 452 unique URLs, with 330 left after exclusions. Reading 330 pages by hand to check whether your brand appears, how competitors are described and whether the facts are current takes days. Claude Code can read the CSV, fetch each page, apply your rules and write structured output, leaving people to review decisions rather than gather data.
Semrush's article uses the Claude desktop app's Code tab and notes a paid Claude plan is needed. The setup rests on three files.
CLAUDE.md: project context loaded every session
Anthropic's Claude Code memory docs explain that CLAUDE.md files are loaded at the start of every session and treated as context, and that the more specific and concise the instructions, the more consistently Claude follows them. Keep it short: what the CSV columns mean, where files live, the output format, and a rule that Claude must quote the page text it relied on for every classification.
brand-facts.md: the approved truth
List positioning, products, current pricing (or "pricing on request" if you do not publish it), integrations, target customers, named competitors and the errors you already know third-party pages repeat. Every correction pitch is checked against this file, so it has to be right. If your positioning has changed recently, our guide to why AI still describes your old brand explains which facts tend to go stale first.
The analysis command
Semrush saves its analysis instructions as analyze-sources.md in .claude/commands/. That still works, but Anthropic's docs now treat the two formats as the same thing:
“Custom commands have been merged into skills. A file at
— Anthropic, Extend Claude with skills — Claude Code Docs.claude/commands/deploy.mdand a skill at.claude/skills/deploy/SKILL.mdboth create/deployand work the same way.”
Either way, you run it by typing /analyze-sources. Arguments typed after the command are available as $ARGUMENTS, which lets you pass a filename. A trimmed example:
---
description: Classify AI-cited pages from a Semrush Sources export
disable-model-invocation: true
---
Read the CSV at $ARGUMENTS. For each URL:
1. Skip our own domain, social profiles, vendor help centres,
directories and knowledge bases. Write review platforms
(G2, Capterra, Trustpilot) to routed-reviews.csv instead.
2. Fetch the page. Record: does our brand appear (yes/no),
competitors named, last-updated date if shown, author/editor
name and contact page if shown.
3. Compare any claim about us with brand-facts.md.
4. Classify: not-mentioned | underrepresented | outdated-or-inaccurate
| negative | no-clear-opportunity | manual-review.
5. Quote the exact sentence(s) that justify the class.
6. Write one specific recommendation (what to add or fix, where).
Output opportunities.csv and manual-review.csv.
Never invent facts not present in the page or brand-facts.md.The six categories come from Semrush's workflow. Semrush also stresses that people must review the output because Claude can make mistakes and may classify borderline pages differently between runs. Requiring a quote for every classification makes that review fast: a reviewer checks the quote against the page instead of re-reading it. For the audit side of this kind of setup, our guide to running an SEO audit with Claude and the Semrush MCP covers connecting Claude directly to Semrush data.
The six page types and what to pitch for each
Each classification calls for a different ask. Sending the same "please add us" email to all of them is the fastest way to burn a target list.
| Classification | What it looks like | What to pitch | Who usually owns it |
|---|---|---|---|
| Not mentioned | Roundup or comparison lists competitors, not you | Inclusion, with a specific reason you fit the page's own criteria | Author or editor of the article |
| Underrepresented | You are listed but with one line while competitors get a paragraph | Additional detail, screenshots, a use case the page lacks | Author or content lead |
| Outdated or inaccurate | Old pricing, discontinued product, wrong feature claim | A single, verifiable correction first | Editor or whoever handles corrections |
| Negative | Critical review or complaint thread | Usually no pitch; fix the product issue or respond publicly where the platform allows | Community, support, PR |
| No clear opportunity | Generic explainer with no brand list | Nothing now; consider original data they might cite later | Not applicable |
| Manual review | Paywalled, ambiguous, or fetch failed | Decide by hand | Your team |
Corrections deserve special attention. They are the easiest yes in outreach because you are helping the editor, not asking a favour. Semrush reports that in one practitioner's last 14 outreach responses, 8 agreed to receive changes without asking for anything in return, using an approach that opens with a single correction and asks permission before sending the full list. That is a small sample from one person, but the principle is sound: lead with one thing that is clearly wrong and easy to fix.
Scoring targets so the first 20 emails go to the right pages
Semrush prioritises by citation impact, prompt relevance, competitor presence, brand gap size, publisher authority and outreach feasibility. Here is one way to turn that into a number. The weights are our suggestion, not a Semrush formula; adjust them to your goals.
| Factor | Score 1 | Score 3 | Score 5 | Weight |
|---|---|---|---|---|
| Citation impact (prompts citing the page) | 1 prompt | 2 to 4 | 5 or more | x3 |
| Prompt relevance | Awareness only | Problem-stage | Comparison or purchase-stage | x2 |
| Competitor presence | None named | One named | Several named, not you | x2 |
| Brand gap | Well covered | Underrepresented | Absent or wrong | x2 |
| Outreach feasibility | No contact, no updates in years | Generic contact form | Named author, recently updated | x1 |
The maximum score is 50. Work top-down.
Worked example: three pages from one export
This is a hypothetical example for an invoicing software brand.
- Page A, "12 best invoicing apps for freelancers" on a mid-sized finance blog: cited for 6 prompts (5 x3 = 15), comparison intent (5 x2 = 10), four competitors named and not us (5 x2 = 10), absent (5 x2 = 10), named author updated this year (5 x1 = 5). Total 50.
- Page B, a software review site's category page: cited for 3 prompts (3 x3 = 9), comparison (10), several competitors (10), our listing shows a pricing plan we retired (5 x2 = 10), but it is a review platform, so it goes to routed-reviews and is handled by updating our own profile (feasibility 5). Total 44, routed not pitched.
- Page C, a general "what is an invoice" explainer: cited for 4 prompts (9), awareness (1 x2 = 2), no brands (1 x2 = 2), not applicable gap (1 x2 = 2), contact form (3). Total 18. Skip.
Page A becomes a pitch this week, Page B a profile fix the same day, and Page C is parked for when you have original data worth offering.
Claude prompts for editor research and first-draft pitches
Once a human has approved the target list, Claude can speed up the next two jobs: understanding each page and drafting a pitch that sounds as though someone read it. Run these per URL, and paste in the page if you are not letting Claude fetch it.
Research prompt:
"Read [URL]. Summarise in 5 bullets: the page's selection
criteria for listed products, how each competitor is described,
the date it was last updated, the author's name and any stated
way to suggest changes. Quote exact text for each point.
Do not guess anything not on the page."
Draft prompt:
"Using the research above and brand-facts.md, draft a pitch to
[author] under 120 words. Open with one specific, verifiable
observation about their page. Make one ask only: [inclusion /
single correction / added detail]. Offer evidence they can
check (public pricing URL, docs page, screenshot). No flattery,
no claims not in brand-facts.md, no mention of AI or rankings."The "no mention of AI" instruction is deliberate. Editors care about whether their page is accurate and useful to readers. Telling them you want to appear in ChatGPT makes the request about you.
Three pitch templates: correction, inclusion and added detail
These are starting points. A reviewer should rewrite at least the first sentence of every email.
Single correction.
Subject: Quick fix on your [topic] page
Hi [Name], your [page title] lists [Brand] at $29 a month. We changed pricing in [month]; the current plans are on [public URL]. Happy to send a couple of other small updates if useful, but only if you want them. Thanks for keeping the list maintained.
Inclusion in a roundup.
Subject: A candidate for your [topic] list
Hi [Name], your list says it focuses on [the page's own criterion, quoted]. [Brand] does [specific fit], which none of the current entries cover. A free trial is at [URL] if you want to test it, and I can share screenshots of [feature]. No problem if it does not fit.
Added detail where you are underrepresented.
Subject: More detail on [Brand] for your comparison
Hi [Name], thanks for including [Brand]. Readers comparing [use case] might want to know [specific capability], which your entries for [competitor] mention. Documentation is at [URL]. I can send a short summary in your format if that saves time.
For subject lines, follow-up timing and reply handling, our guide to link building outreach emails that get replies goes deeper.
Where AI citation outreach crosses the line
Because the goal is a mention rather than a link, some teams assume search engine rules no longer apply. They do whenever a link is involved. Google's spam policies are explicit:
“Link spam is the practice of creating links to or from a site primarily for the purpose of manipulating search rankings.”
— Google Search Central, Spam policies for Google web search
The same page lists exchanging money, goods or services for links, and sending someone a product in exchange for writing about it and including a link, as examples. Paid placements are fine if the links are qualified with rel="sponsored" or rel="nofollow". Beyond Google's rules, avoid:
- Paying for undisclosed inclusion in "best of" lists. It risks link spam if a followed link is involved and may breach advertising disclosure rules in your market.
- Fake or incentivised reviews on review platforms to change how a page describes you.
- Mass AI-generated pitches sent without human review. They read as spam and can damage the relationships that make outreach work.
- Asking for competitors to be removed or disparaged. Ask for accuracy about your brand only.
- Publishing fake "independent" comparison sites. AI engines may cite them for a while; editors and users notice.
The sustainable version is the one that also earns links and coverage, covered in our guide to digital PR that earns mentions LLMs trust.
Tracking whether outreach moved the AI answers
A reply is not a result. The result is a changed page, and then a changed answer. Keep a tracker with these columns, then re-export the Sources report monthly to check whether each page is still cited and whether your brand now appears.
| Column | Example value | Why it matters |
|---|---|---|
| URL and prompts citing it | Page A; 6 prompts | Lets you check the right answers later |
| Classification and score | Not mentioned; 50 | Shows whether high scores produce more wins |
| Ask sent and date | Inclusion; 6 Oct | Time-to-update benchmarks |
| Outcome | Added 21 Oct | The real win metric |
| Brand mention in answers after 30 and 60 days | Named in 2 of 6 prompts on ChatGPT | Connects page change to AI visibility |
| Page still cited? | Yes | Cited sources churn; a win on a dropped page has less value |
Expect churn. Semrush's study of more than 100 million citations found Reddit fell from close to 60% of ChatGPT responses in early August 2025 to about 10% by mid-September, and Wikipedia from roughly 55% to under 20%. Citation sets move, so measure over months rather than weeks and report on the trend. Our guides to link building KPIs and to building an AI visibility dashboard in Semrush show how to put these numbers in front of stakeholders. If the answers you are trying to change are actively steering buyers away, see what we found about AI chatbots talking buyers out of purchases.
Questions SEO teams ask about AI citation outreach
Is AI citation outreach just link building with a new name?
The mechanics overlap, but the target list and success measure differ. You choose pages because AI engines cite them for your prompts, and success is an accurate mention that shows up in answers, whether or not a followed link is involved.
How many prompts do I need to find useful targets?
Twenty-five well-chosen prompts, the AI Visibility Toolkit's limit, is enough to surface a few hundred cited URLs. Weight them toward comparison and problem prompts, where brands are named most often.
Can Claude send the outreach emails for me?
It can draft them, but a person should review and send every one. Semrush says human review of Claude's classifications is required, and unreviewed AI pitches damage editor relationships.
How long until changes show up in AI answers?
There is no published timeline. Re-export monthly and compare mention rates at 30 and 60 days after a page changes; citation sets also shift on their own, so judge trends.
Should I pitch review sites like G2 or Capterra?
Usually not with an editorial pitch. Route them separately and update your own profile, encourage genuine reviews, and respond to feedback within each platform's rules.
Where to start with your first 30 cited pages
Set up 25 prompts weighted to comparison and problem intent, wait a day for data, export the Sources report, and run a Claude triage with a brand-facts file you trust. Score the opportunities, have a person review the top 30, and send five single-correction emails before anything else. Corrections build replies and credibility you can use for harder inclusion requests. Start the tracker on day one, because without it you will not know which wins changed the answers.
If you would rather have a team run this for you, from auditing which pages AI already cites to outreach and reporting, our SEO services and content writing teams can help. Get a quote or request a custom quote for an AI citation outreach programme.
Further reading on ai citation outreach
- How to Run an SEO Audit With Claude and the Semrush MCP
- Building an AI Visibility Dashboard With Semrush
- Google AI Overviews Now Show for Most Major Brands: What to Do
Sources for this AI citation outreach workflow
- Semrush: AI citation outreach with Semrush and Claude (21 September 2026)
- Semrush: Why 62% of AI citations don't lead to brand mentions [Study] (9 June 2026)
- Semrush: How to track your AI citations and sources with Semrush (7 October 2026)
- Semrush Knowledge Base: Prompt Tracking
- Semrush: The Most-Cited Domains in AI: A 3-Month Study (10 November 2025)
- Semrush: Do Backlinks Still Matter in AI Search? Insights from 1,000 Domains (16 October 2025)
- Anthropic: Extend Claude with skills — Claude Code Docs
- Anthropic: How Claude remembers your project — Claude Code Docs
- Google Search Central: Spam policies for Google web search
