AI Citation Outreach: How to Turn Citation Data Into a Workflow Using Similarweb MCP

AI citation outreach, the off-page side of AI search optimization, means using citation data to find the third-party pages AI answers rely on, then asking the publisher for a specific, accurate change. With Similarweb MCP, Claude, Manus, and many other AI tools can pull, sort, and score that data for you. Once you trust the data, you just review the results and send the pitches.

The reason to act is simple: most of what AI cites about a SaaS brand lives on other people’s pages, so optimizing your own site for AI search covers only a small share of the picture. Across 15 SaaS brand ecosystems, third-party sources made up 84% to 93% of citation weight in ChatGPT and Google AI Mode, according to Aleyda Solis’s benchmark. In B2B SaaS, I treat third-party pages as the main channel for AI visibility, not an extra.

Citations are also becoming more common, which matters to anyone practicing AI SEO. The share of ChatGPT answers that include web citations grew more than 5X between June 2025 and May 2026, to 6.8% overall and 6.6% in Technology, according to Similarweb’s 2026 Generative AI Landscape report.

This article covers five types of outreach opportunities in your citation data, a six-step weekly workflow called the Citation Gap Loop that you can schedule to run automatically, a score for ranking targets, and the pitches and tracker that finish the job. The workflow section has paste-ready prompts for the data steps, which you can run in Claude with Similarweb MCP. The examples use one B2B brand tracked on ChatGPT and Google AI Mode, compared week over week.

Key takeaways:

• Start from the pages AI already cites, not from a list of top publishers.
• Sort your citation export into five opportunity types, because each one needs a different request.
• Pull each LLM separately and compare week over week. In the example, neither ChatGPT nor Google AI Mode cited the same page.
• Score and verify each page, pitch one specific edit, and track the result every week.

What is AI citation outreach, and how is it different from link building?

AI citation outreach is the part of AI search optimization that happens off your site. It’s a slightly different tactic from link building in where targets come from (cited pages in your citation data, not authority metrics) and in how it is judged (mentions in tracked prompts, not links acquired).

Writing the email is easy. The hard part is choosing which of several hundred cited pages deserves one.

Aleyda Solis’s benchmark makes a related point. One accurate placement on a comparison page AI often reuses is worth more than broad coverage on many low-impact pages. Correcting how you appear on that page matters as much as appearing on it.

The right targets also depend on the large language model (LLM). In the same benchmark, technology publications made up 7.7% of ChatGPT’s external evidence but 0.5% of Google AI Mode’s. Review platforms made up 6.4% of ChatGPT’s external evidence and 2.8% of Google AI Mode’s.

To show the same split, I chose a campaign in Similarweb’s AI Brand Visibility tool that tracks monday.com on ChatGPT and Google AI Mode in the US. These are its citation counts by domain in the week of October 14 to 20, 2026:

A dash means the domain was not among LLM’s 30 most-cited domains that week.

If you pull data for one LLM and pitch as if it covered all of them, you may spend effort on pages the other platforms ignore. I pull each LLM separately, because in this campaign a combined number hid every target. Perplexity covered only about 7 responses that week, which is too few to read, so the examples use ChatGPT and Google AI Mode.

Some teams call the overlap citation link building or AI link building: earning a link on a page that AI already cites. A link can result from the request, but it is not the goal. Do not confuse it with local business citation building.

Run both activities, but keep separate trackers. Link building supports classic SEO, while AI citation outreach supports AI search optimization, so the two serve different goals. The same publisher can appear in both, and one tracker would hide which program is working. Similarweb’s backlink analytics covers the link building side.

If you already run link building outreach, you can reuse your contact process and templates. Only the target list changes.

Before your next outreach batch, tag each target as coming from citation data or from an authority list. The success metric above depends on the difference between mentions and citations. For the details, see AI mentions vs. AI citations.

What are the 5 types of AI citation outreach opportunities?

AI citation outreach opportunities fall into five types: pages cited where competitors appear and you do not, citations you lost since last week, answers that name you without citing you, cited pages that describe you wrongly, and rising sources worth reaching early. Each type needs a different request, so sort the export before you pitch anyone.

1. Competitor-cited gap

• How to detect it: A prompt that names competitors, omits you, and cites a third-party page.
• What to ask for: Add your entry.
• Typical page: Roundup, ranking, or comparison.
• Effort: Medium.

Take the prompt “What are the differences between collaboration and project management tools?” It ran once a day on ChatGPT and on Google AI Mode in the AI Brand Visibility tool between October 14 and 20, 2026. The brand tracked in that campaign is monday.com.

This is competitor citation analysis at the page level: you compare who is cited and who is named, then act on the difference.

The two LLMs gave very different answers:

No cited page appears in both lists. This is why you pull and score each LLM separately. Google AI Mode already names the brand in this prompt more often than it names competitors, so the gap there is small. In ChatGPT, the brand was named on 1 of 7 days, and it cited the target pages. For the full method behind this gap, see the guide to citation gap analysis. I start with this type: the request is one entry on a page that already exists.

2. Lost citation

• How to detect it: Same URL, latest 7 days vs. prior 7 days. The prior week had at least 10 citations, and the drop holds on the next pull.
• What to ask for: Refresh or restore your entry or data.
• Typical page: Any page that used to cite you.
• Effort: Medium.

Check the URL and the threshold, not just the domain. The domain-level pull flags itechguides.com: its citations across the tracked LLMs fell from 19 in the week of October 7 to 13 to 14 in the week of October 14 to 20, a 26.3% drop. That is under the 30% threshold, so it goes on a watch list instead of the outreach queue.

The same drop rule caught smartsheet.com in ChatGPT, which fell from 37 citations to 20 or fewer. Smartsheet is a direct competitor, so I exclude it before anyone pitches it.

When a real loss passes the threshold on a pitchable page, check the URL level and confirm the drop on the next weekly pull. Match the request to the page type. A roundup needs an entry, and a reporter-led feature needs an expert quote or a data point for the next update.

3. Mentioned, not cited

• How to detect it: The answer names you, but none of your URLs are cited.
• What to ask for: Nothing from the publisher. Build a page that earns the citation.
• Typical page: Competitor-hosted or third-party pages.
• Effort: High (content task).

Treat this as a content task first. On October 17, ChatGPT named Monday in its answer to the prompt above and cited pages from Computerworld, Kanbanchi, Notion, and Wrike. On October 20, Google AI Mode named Monday and cited Quire, Quizlet, Unito, and Atlassian pages.

In both cases, the brand was named, but none of the brand’s URLs were cited.

A publisher cannot fix that. A page on your site that earns the citation can, and building it is on-site AI SEO work. The AI citation analysis guide explains how to build one. I treat these as a content backlog first, then an outreach list.

4. Misrepresented page

• How to detect it: The page is cited and names you, but a fact is wrong or out of date.
• What to ask for: One correction, with proof.
• Typical page: FAQ, pricing mention, or comparison row.
• Effort: Low.

Open the page and count. The Unito blog post lists five tools and names monday.com once, in passing, inside the ClickUp entry. The Computerworld feature does not name the brand at all.

Neither page states a wrong fact, so the correction request did not apply this week. The same check finds wrong facts when they exist. The verification table later in this article covers both pages.

Pricing claims are the simplest to check. The Digital Project Manager’s collaboration tools list, which ChatGPT cited twice that week, says monday.com has a free plan and starts at $9 per user per month, billed annually. Both claims match monday.com’s pricing, so no correction is needed. A page that listed different prices would be flagged.

5. Rising source

• How to detect it: A URL or domain gaining citations two weeks in a row.
• What to ask for: Get in before the page is crowded.
• Typical page: Fast-growing roundups.
• Effort: Low to medium.

In the week pair above, these domains gained citations:

One week of gains is a watch item, not a trend. Add these domains to a list and confirm on the next pull. Note which LLM each gain is in. techradar.com and g2.com are rising in ChatGPT, while Google AI Mode leans on reddit.com and zapier.com.

What does not belong on the outreach list?

Three source types look like targets but are not.

These pages still show what content to build. They just do not belong in the outreach queue. A government site will not add you to its list, no matter how polite the email.

How to automate AI citation outreach with Similarweb MCP: the Citation Gap Loop

The Citation Gap Loop has six steps. Similarweb MCP and Claude can do most of the work: pulling cited URLs, sorting them, reading the pages, scoring targets, finding contacts, and drafting the email. Run it on a weekly schedule to refresh your report without anyone pasting a prompt. You review the results and send the pitches.

Every step uses a week-over-week window. Citation counts shift week to week, so a single week tells you little.

The exact Similarweb MCP tool names, which you will need for the prompts below:

How do you set up Similarweb MCP for citation outreach?

Connect the Similarweb connector in Claude. Similarweb’s Claude integration shows how, and the GEO keyword research automation guide shows the same setup for keywords.

Then ask Claude to list your AI Brand Visibility campaigns with get-gen-ai-campaigns. It returns each campaign’s ID, tracked domains, brands, competitors, LLMs, and date range. Use the campaign ID in every pull.

Five setup tips that save time:

• Tell each citation call which LLMs to check (chatgpt, google_ai_mode, perplexity). If you leave this out, you get one combined total for all LLMs and cannot see how they differ.
• Visibility calls need exact dates in the format YYYY-MM-DD. Prompt-level data covers at most 90 days at a time, so a one-week pull is well within the limit.
• The domain filter matches one exact domain and does not include subdomains. Filter support.monday.com and monday.com separately (16 vs 34 citations in the week of October 14 to 20).
• Each campaign has one country. You cannot set a different country on a citation call.
• More detail costs more credits: In my pulls, domain-level calls cost 1 to 2 credits, a call that lists the cited pages of one domain costs 1, and each prompt-level row costs about 5. Pull prompt-level data only for pages that pass the sorting step.

Which prompts should the pull cover?

The pull only finds pages for the prompts you track, so the prompt set decides your whole target list. Cover two kinds: awareness prompts, where buyers explore a problem, and evaluation prompts, where they compare options, like the collaboration-tools prompt in this example. Evaluation prompts matter most for outreach, because that is where brands get compared and where adding your entry changes the answer.

Leave out prompts that contain your own brand name. They make your brand look more visible than it is and hide the pages you need to be on. Keep the set fixed for a few weeks, or week-over-week changes will reflect the prompt set and not the citations.

To find candidate prompts, see Similarweb’s guide to GEO keyword research. To decide which ones earn a permanent slot, see How To Choose Which AI Prompts Are Worth Tracking.

Which brand facts should Claude check pages against?

The Verify step can only flag a wrong or outdated claim if it knows what is right. Keep one short document of approved facts and give it to Claude with every run. Include your positioning, current pricing and plan names, the products and features you want pages to get right, your direct competitors, and any wrong claims third-party pages often repeat.

Keep it to one page. A long file slows the run and makes it harder for Claude to tell which fact matters. Update it whenever pricing or plan names change, because a stale file produces false corrections. In the scoring prompt, [BRAND_FACTS] stands for that file.

What does the weekly pull look like?

This is the instruction I paste into Claude. Add your campaign ID and dates.

For campaign [CAMPAIGN_ID], call get-citations-domains-agg once for each LLM in [LLM_LIST] (llms filter), for [CURRENT_WEEK_START] to [CURRENT_WEEK_END] and again for [PREVIOUS_WEEK_START] to [PREVIOUS_WEEK_END], sorted by citation_count, limit 100. Return one table per LLM with domain, source_category, citations in each week, absolute change, and percent change. Exclude source_category “Direct Competitors”. Flag any domain whose prior-week count was 10 or more and fell by 30% or more.

In this campaign, each prompt runs once a day per LLM. A page cited for one prompt can therefore reach at most 7 citations a week in each LLM, or 14 across ChatGPT and Google AI Mode.

Read citation_count together with prompt_count. Otherwise, a page cited every day for one prompt will look weaker than a page cited twice across ten prompts.

Steps 2 to 4 use two more prompts. The next section defines the Outreach Priority Score and the feasibility scale they mention. The first pulls URL-level data for each flagged domain and sorts it:

For each flagged domain, call get-gen-ai-campaign-analysis-citations-aggregated with the domain filter for both weeks, limit 25. Return URL, citations and prompt_count in each week. Tag each URL: Lost citation (prior week 10 or more, down 30% or more), Rising source (up two weeks running), or Watch. Exclude government, academic, and vendor documentation domains.

The second scores the pages that remain:

For the top 25 tagged URLs, call get-gen-ai-campaign-analysis-prompts for the current week with metrics prompt, brands_mentioned, citations, once for each LLM. Count how many days each URL was cited and how many of those days my brand was named. Call get-websites-traffic-and-engagement weekly for each domain.

Open each page, count my brand’s appearances, classify the page, and set source feasibility from the scale. Compare each page’s claims about my brand with [BRAND_FACTS] and quote any claim that differs. List any page you could not open. Compute the Outreach Priority Score with weights 30, 20, 25, 15, 10.

For each target, set Link to add from the My Domain pages for the same topic, using get-gen-ai-campaign-analysis-citations-aggregated on the tracked domain for both weeks. Prefer a page cited this week that also appears in get-keywords-latest-agg for the topic’s main keyword, with clicks, position, and top_url. A page cited last week but not this week is also eligible, even if it does not rank. If no page qualifies, use the closest one and write “unverified” next to it.

For step 6, run the first prompt again, moving both windows forward by 7 days. Then call get-gen-ai-campaign-analysis-visibility-aggregated to log brand visibility for the same week.

No MCP connector? Run the same steps by hand:

1. Export the cited URLs and domains per prompt from any citation tracker, for the latest 7 days and the prior 7 days, one LLM at a time.
2. Paste them into a spreadsheet with one row per cited page, using the columns in the tracker table below, and fill in the days cited and the days your brand was named.
3. Open each top page yourself, count your brand’s appearances, and set the feasibility score.
4. Calculate the week-over-week change, the gap, and the Outreach Priority Score with the formulas in the next section.

How do you run the weekly pull automatically?

Once the prompts work, schedule them. The pull, sort, and score steps need no judgment, so they can run without you and update your report each week. Claude can also read the top pages and classify them. You review the results and send the pitches.

You have three options:

• Claude scheduled tasks: Save the prompts as a scheduled task that runs every week. Desktop scheduled tasks are persistent, but they only run while the app is open and your computer is awake. Cloud tasks run without your machine. Tasks created with /loop expire after 7 days, so do not use them for a permanent weekly run.
• Manus: Connect Similarweb MCP by following Similarweb’s Manus integration guide. Then create a task in Manus Scheduled Tasks, which supports weekly schedules and keeps a run history.
• Claude Code: Create a routine in Claude Code. Like Manus, it runs in the cloud on a weekly schedule, calls your connectors (including Similarweb), and keeps a run history. It cannot open files on your computer, so keep your tracker data in a private GitHub repository and leave out contact details.

Other tools that support MCP and scheduling can run the same prompts.

How to set up the scheduled run:

1. Run the prompts by hand first and check the numbers.
2. Make sure the Similarweb MCP connection is available to the app that runs the task.
3. Choose a Desktop task if your tracker file is on your computer. Cloud tasks start from a fresh clone and cannot see your local files.
4. After creating a Desktop task, click Run now and approve the tool prompts, or the run will stall waiting for approval.
5. Ask the task to pull both weeks for each LLM, sort the pages, read the top ones, score them, and save the output to your tracker data.
6. Ask it to rebuild the HTML dashboard from the updated tracker data.
7. Ask it to list new lost-citation and rising-source flags at the top of the report.
8. Review the results before pitching. Do not schedule the pitch step. I keep the pitch step manual, because a wrong fact in an email costs more than the minutes it saves.

How to choose which cited pages deserve a pitch

Which sources can you pitch?

Filter by source category, then verify what is left. The categories in the export are broad, so use them as a first cut, not a final answer.

One caution. In this campaign, vendor blogs such as unito.io and kanbanchi.com carry the same Business Services label as vendor documentation such as learn.microsoft.com. The label alone does not tell you what to do, so open the page.

How do you verify a page before pitching?

Open the page, count how many times your brand appears, and classify what you find. The five labels are absent, underrepresented, outdated or inaccurate, negative, and no opportunity.

Claude can do this first pass, and you should review the results before sending any pitch. A pitch with a wrong fact is likely to be ignored. I do not send a pitch for an unverified page. One wrong fact costs me the editor’s trust.

Two pages from the collaboration prompt show the range:

Why shouldn’t you filter targets by publisher traffic?

Traffic and citations do not move together. Similarweb data for these pages shows:

Computerworld has about four times Kanbanchi’s traffic and half the citations. The model did not read the traffic report.

A list that starts with traffic or domain authority would rank Computerworld first and skip the page ChatGPT cited most. I keep publisher reach in the score, but I weight it low.

Similarweb’s 2026 report shows the same gap at scale. In May 2026, 65% of the URLs ChatGPT cited sat two or three folders deep, while 58.8% of ChatGPT referral traffic landed on homepages (US, desktop). Aleyda Solis, quoted in the report, describes it as “a clear disconnect” between the pages ChatGPT cites and those receiving referral traffic. Score cited pages and traffic pages separately.

Does the loop change for ecommerce brands?

The steps stay the same. The source mix changes. Similarweb’s 2026 report found that:

• Beauty conversations on ChatGPT lean heavily on ecommerce brands.
• Travel conversations lean on reviews and user-generated content.
• Finance conversations lean on specialist financial publishers.

Retail answers also carry citations more often than technology answers: 13.5% against 6.6% of ChatGPT answers in May 2026. A retail brand therefore has more cited pages to work with.

Run the pull for each category and rewrite the feasibility scale for your sources. For an ecommerce brand, replace vendor blogs and publisher roundups with review platforms and retailer-adjacent roundups in the example above.

What does the weekly routine look like?

Without MCP or a scheduled task, you do the Monday and Wednesday work yourself with a manual export. The routine is the same, but it takes more of your time.

Rank pages by Outreach Priority Score, because authority metrics pick the wrong targets

The Outreach Priority Score ranks each cited page from 0 to 100 using five inputs: citation weight, prompt reach, competitor gap, source feasibility, and publisher reach. The weights are 30, 20, 25, 15, and 10.

I put proof of influence ahead of publisher size because how often AI already cites a page counts for far more than how big the publisher’s site is.

Here is the score applied to three pages cited for the collaboration prompt in the week of October 14 to 20:

How to read this table:

• Scores are relative to the set you scored. A high score means high priority in that set, but it doesn’t guarantee anything.
• Kanbanchi ranks first on a provisional feasibility of 0.6. It is a project management vendor, so the page may be written to promote its own product. Verification can lower the score.
• Prompt reach does not separate these pages, because all three were cited for one prompt. This is why you score the full export.
• The weights are my starting point, not a validated model. Adjust them against your own results after 8-12 weeks.

Two rules apply on top of the score:

1. A lost citation goes to the front of the queue, because the page once included you.
2. No pitch goes out before the page is verified, because a score based on an unchecked page is unreliable.

A good citation outreach pitch asks for one edit, and the tracker shows whether it landed

An effective citation outreach pitch asks for one specific, verifiable edit, such as adding an entry, correcting a fact, or refreshing a price, and offers the supporting detail ready to paste. The tracker then records the request, the date it went live, and the first weekly pull in which the page and the brand appear together.

How do you find the right contact?

Use Similarweb Sales Intelligence contacts first. See the contact information page for what the data covers. Claude can query it through two tools in Similarweb MCP:

• Contact search: search by the publisher’s domain and a job title such as “editor,” with has_email set to true. It returns each contact’s title, seniority, department, and an accuracy score.
• Contact enrichment: enter a full name and the domain to get that contact’s email and details. The bulk version does the same for up to 25 contacts at once.

In a test on computerworld.com, a search for editor titles returned contacts with available emails and accuracy scores of 90-92.

Revealing a contact uses credits in Similarweb Sales Intelligence, so reveal only the contacts you plan to pitch.

If the search returns no one useful, fall back to the page. The byline or author page usually names the writer or editor, and Claude reads it when it opens the page. Claude can then look for a contact or partnerships page on the same site.

Pick the person whose job matches the page: the writer for a feature, the editor for a roundup, a partnerships or marketing contact for a vendor blog. Check the accuracy score before you send, and follow your company’s rules for using contact data in email outreach.

What should the pitch say?

I lead with the edit, not with AI. A clear change to one section is easier for an editor to approve than a general request.

The citation data explains why you chose the page. I keep it out of the email.

Every pitch has the same seven parts:

1. Subject line: names the page or the change, in about eight words or fewer.
2. Opening: one specific detail from the page, so the editor knows you read it.
3. Who you are: one line with your name, role, and company.
4. Request: one edit, in one named section.
5. Ready-to-use text: the draft entry, corrected line, or figure, pasted into the email, so the editor does not need a second round of messages.
6. Proof: one link to a primary source, such as your pricing page.
7. Exit: a line saying it is fine if the edit does not fit.

Send plain text with no attachments, and keep each email under about 130 words, not counting the pasted text. Claude fills the placeholders from the page and your brand facts. You review the opening line before sending.

Each email needs three tracker fields:

• Contact: the name of the person the email is for, from the contact search.
• URL: the cited page you want a mention in, from the citation data.
• Link to add: your own page for the same topic (source category My Domain). Prefer a page AI cites now that also ranks for the topic’s main keyword. A page cited in the prior pull that lost its citations also counts, even if it does not rank. If nothing fits, use the closest page and mark it unverified.

Owner (the sender’s name and title, such as “Dana Lee, head of content”) and Brand stay the same for every email. Claude reads the page at the URL and fills in everything else.

Here are three templates, one for each opportunity type, and a follow-up for all of them.

Pitch template: add your entry (competitor-cited gap)

Subject: Draft entry for [page title]

Hi [Contact],

I’m [Owner] at [Brand]. I read [page title], especially [one specific detail].

It compares [number] [products] and was last reviewed on [date]. [Brand] is not in it.

Could you add it to the [section] section, linked to [Link to add]? Here is a draft entry in your format, ready to paste:

[Draft entry, 60 to 80 words: what the brand is and who it suits]

If [Brand] does not fit your criteria, no problem.

 

Thanks,

[Owner]

[Brand]

Pitch template: correct one fact (misrepresented page)

Subject: One correction on [page title]

Hi [Contact],

I’m [Owner] at [Brand]. Your [page title] says “[exact claim]” in the [section] section.

Our [pricing or product] page shows [correct fact] ([Link to add]), last updated on [date]. Here is a corrected line you can use: “[Corrected sentence, in the page’s own style]”

Thanks for keeping the page accurate for your readers. I also have [one related fact] for the same answer, and I can send it if it is useful.

 

Best,

[Owner]

[Brand]

Pitch template: refresh an outdated page (lost citation)

Subject: New data for your [article title]

Hi [Contact],

I’m [Owner] at [Brand]. I read your piece on [topic], especially the point that [specific point].

Since it ran, [one new data point] ([Link to add]) shows [what changed]. If you update the article, I can send the full figure, the method, and a two-line quote from [expert, title] that you can use or edit.

Would that help with a refresh? If it does not fit, no problem.

 

Thanks,

[Owner]

[Brand]

Follow-up template (same thread, any type):

Subject: Re: [original subject]

Hi [Contact],

Following up on my note about [the one edit, in a few words]. The [draft entry, corrected line, or figure] is in my first email, and I can resend it in whatever format suits you. If someone else looks after this page, could you point me to them?

 

Thanks,

[Owner]

[Brand]

Follow up once, four days after the first email. Close the item after the next weekly pull. Do not add false urgency, and do not put a list of five requests in a first email.

What can you offer in return?

Decide before outreach starts. Common options are:

• A link from a page on your site where it fits naturally
• Free or discounted product access
• A data contribution
• A paid placement, within a budget the outreach owner sets

Log every offer next to the target in the tracker. Later, you can see which requests cost money and which didn’t. I start with the first three and treat a paid placement as the last option.

How do you track results and build the dashboard?

Use one row per cited page and update it every week. The core columns are below. Together they work as an AI citation tracker for outreach targets.

The dashboard needs the five numbers that show whether outreach is working:

• Targets by pipeline stage
• Placements live
• Brand visibility on the prompt set, against your baseline
• Share of your top 10 scored targets contacted
• Median days from pitch to live

You do not need to build the dashboard by hand. Claude can build a designed HTML dashboard that opens with the data, ends with ranked actions, and updates weekly. Run this after the pull and scoring prompts, in the same conversation, so Claude has this week’s results:

Build a one-page HTML dashboard for AI citation outreach on campaign [CAMPAIGN_ID], using this week’s pull results and the week before for comparison. Make it a single self-contained HTML file I can open in a browser. Use dark navy (#092540) for text, bright blue (#195AFE) as the main accent, orange (#FF7A1A) for highlights and actions, and light grey (#F7F7F8) for the page background. Use DM Sans or Arial, and support dark mode.

Open with five key metrics: targets by pipeline stage, placements live, brand visibility on the prompt set against the baseline, the share of the top 10 scored targets contacted, and the median days from pitch to live. Color the numbers that need attention. Then show the ranked targets with status chips, brand visibility by LLM, and week-over-week citation movement.

Finish with a “What to do this week” panel of 3 to 5 ranked actions, each tagged Pitch, Check first, or Watch. Base them on these rules: pitch the highest-scored verified pages first, check unverified pages before pitching, flag lost citations (prior week 10 or more, down 30% or more, confirmed on the next pull), and put one-week gains on a watch list.

Save the data behind the dashboard in a tracker spreadsheet: each week’s pull results and my outreach log (pitches sent, replies, and live dates).
Give the tracker and the HTML file clear names. Each Monday, after the scheduled pull finishes, add the new week’s rows to the tracker and rebuild the dashboard from the same HTML file, replacing only the data.

The dashboard opens with the data and ends with the actions, so you see the numbers before deciding what to do. The key metrics that need attention are highlighted in color. Replace [CAMPAIGN_ID] with your campaign ID. Claude creates the tracker and the HTML file, so you don’t need to name either one.

This is what it would look like:

Access the live HTML here: AI Citation Outreach Dashboard Example

The refresh instruction belongs in the prompt, because an HTML file cannot call Similarweb MCP on its own. The scheduled task pulls the data, writes it to the tracker file, and rebuilds the page. The dashboard is as fresh as your last run.

This example covers one brand, one campaign, and one pair of weeks, and I haven’t sent the pitches in this workflow, so I have no outcome data. The case for placements rests on the third-party citation data above, not on measured results. Treat the scoring weights and thresholds as starting points to test.

I do not have a verified typical lag, so I will not quote one. Publishers update on their own schedule, and a model reflects a page only when it next retrieves it. The tracker measures this: days from the live date to the first weekly pull where the page is cited with your brand named.

Report the change across the whole prompt set. One page moving one prompt is not a result.

How to start AI citation outreach this week

AI citation outreach works best as a weekly routine. Pull the week for each LLM, sort the opportunities, check the pages, score them, send one edit per email, and read the next pull.

Your citation data already shows which pages AI relies on, so you do not have to guess which publishers matter. It also gives your wider AI search optimization program a weekly number to report: brand visibility on your tracked prompts.

The collaboration prompt shows why each LLM needs its own pull: a single-LLM pull would have shown only half the picture.

To run the first pull on your own domain, use Similarweb MCP and Similarweb’s AI citation analytics. They return the cited URLs and domains per prompt that this workflow starts from. Then review the results and send the pitches.

FAQ

What is citation outreach, and is it the same as citation link building?

Citation outreach, also written AI citation outreach, uses citation data to find the third-party pages AI answers cite, then asks publishers for a specific, accurate change. Citation link building is the narrower version, where the goal is a link on such a page. Citation outreach is one tactic within AI search optimization and is judged by brand visibility on tracked prompts, so a link is optional.

How do I run AI citation outreach with Similarweb MCP?

Connect the Similarweb connector in Claude and list your campaigns. Run the cited domains and cited URLs reports for the latest 7 days and the prior 7 days, once per LLM. Add the daily answers and website traffic reports, then let Claude read the top pages. Save the prompts as a weekly scheduled task so the report refreshes itself.

Does AI citation outreach replace link building?

No. Link building starts from the pages you want to rank and measures links and traffic, while citation outreach starts from the pages AI cites and measures brand visibility on tracked prompts. Aleyda Solis’s benchmark found third-party sources carry 84% to 93% of citation weight for SaaS brands, so both programs often reach the same publishers.

How do I find which pages AI cites for my category?

Track a fixed set of prompts, then pull the cited URLs and domains for the latest 7 days and the prior 7 days, one LLM at a time. Rank them by citation count and prompt count, remove sources you cannot pitch, and read the rest. In the example campaign, ChatGPT and Google AI Mode cited entirely different pages for the same prompt.

How do I find citations I lost?

Compare the same URL across two weekly windows and flag pages that had at least 10 citations in the prior week and fell by 30% or more. Confirm the drop on the next pull before acting. In the example campaign, itechguides.com fell from 19 to 14 citations (-26.3%), so it stayed on the watch list.

Why should I pull each LLM separately?

Each LLM cites different pages, so combining them hides your targets. For the same prompt in the example campaign, ChatGPT named the brand on 1 of 7 days and Google AI Mode on 4 of 7, and no cited page appeared in both lists. Score each LLM on its own, then merge the targets.

How do I measure whether citation outreach worked?

Track brand visibility on your prompt set every week, and log the first week a pitched page is cited with your brand named. Report the change across the whole prompt set against a baseline, not one page or one prompt. The tracker records the pitch date, the live date, and that first pull.

How long does AI citation outreach take to work?

There is no reliable benchmark, because it depends on how fast the publisher updates the page and when the model next retrieves it. Measure it instead. Log the date each edit goes live and the first weekly pull in which the page is cited with your brand named, then report the median across placements.

Should I pitch competitors’ blogs and comparison pages?

Not for listings. A page hosted by a direct competitor is written to favor that competitor, so an inclusion request is unlikely to go anywhere. Treat it as a signal of what content to build on your own site. Independent roundups and review sites are better targets, though you may want to correct a competitor’s site if it states a wrong fact about you.