Salesforce AI citation outreach: find the pages AI engines cite, then check your CRM

Semrush defines AI citation outreach as the process of finding pages AI answers cite and getting your brand included on those pages. It is not the same as link building: the goal is for a page the engines already trust to mention you accurately, not for a backlink.

It matters because citing and naming are different events. A Semrush study with Kevin Indig, quoted in its guide, found that 62% of AI citations do not result in a brand mention, which it calls ghost citations. Similarweb's guide sorts the targets into five types: competitor-cited gaps, lost citations, mentioned but not cited, misrepresented pages and rising sources.

The distinction between being cited and being named shows up in practice. A news site publisher on Reddit wrote: "Cited pages sit at roughly 700. Brand mentions sit at five." Citation outreach starts from that gap, so the list of pages worth contacting should be tied to a question where you want to be named.

No, and the search results for this query mix the two up. Salesforce's citations feature links an Agentforce agent's answer to the sources behind it, such as knowledge articles, PDF files and external web pages, so a user can check the answer. The developer guide says platform-generated citations are only available for agents created after May 26, 2025.

That is your own agent citing its sources. Outreach is the reverse job: a public AI engine cites someone else's page, and you want to be on it. Nothing in Salesforce measures what ChatGPT or Perplexity cite, so the measurement has to come from somewhere else.

The Salesforce connector is the relationship side. Its listing shows four tools: discover, describe, dispatch and dispatch_readonly. Salesforce's reference says discover searches the library of available operations, describe returns the specification, dispatch runs one and dispatch_readonly runs read-only operations that never change data. It does not name Salesforce account fields, so Claude finds them with discover first.

Linkeddit is the evidence side, one plan at $60 a month. answer_gaps_get lists the open questions where an engine recommends a competitor over you, and answer_sources_fetch reads a cited page and stores it in your Answer Radar evidence store, so Claude can check what it says about you. Because it stores data, the prompt asks for your approval before it runs.

On Linkeddit's test project, plausible.io, a read-only answer_gaps_get call returned one open gap: 'Not represented for: Umami vs Matomo for self-hosted analytics', rank 1, with an evidence score of 22 and intent score 0.6. The list was short, which is honest: a gap is a question where you were absent, not a guess.

For each gap, Claude reads the cited pages, then asks Salesforce through dispatch_readonly whether each publisher already exists as an account, contact or lead, and who owns it. A publisher you already know needs a warm introduction from its owner. A new one needs a cold pitch. A competitor's own site is rarely worth pitching, which Semrush lists as a feasibility factor.

Linkeddit returns the cited page and its domain, not a contact's email address. Claude finds the right person from the page itself, such as a byline or contact page, or from a Salesforce contact that already exists. It never guesses an address.

Semrush ranks targets by citation impact, prompt relevance, competitor presence, the size of the brand gap, publisher authority and outreach feasibility. A good pitch asks for one specific edit: adding you to a comparison, or correcting a fact. Claude drafts the plan and the wording, and stops before anything is logged to Salesforce. The log step uses dispatch, which can create records, so approve it page by page.

Salesforce says every Hosted MCP transaction runs as the authenticated user through an External Client App, so object permissions, field-level security, sharing rules and permission sets all apply. If you cannot do something in Salesforce, your agent cannot do it through the server. Use a login that sees only the account, contact and lead objects this job needs.

Salesforce also advises setting tool-level restrictions in your client, so a tool that changes configuration or data asks for your approval first. Because the server is in Beta and its operations grow with each release, ask Claude to run discover and read the result before relying on any operation by name. Read the draft plan, check each publisher and each claim, and only then approve a log entry.