Check this page with an assistantOpens a chat asking it to summarise this article and name the evidence behind each claim.
Claude opens with the prompt on your clipboard: Anthropic does not support prefilled prompts on the web, and we would rather copy it than ship a button that drops it.
The Finding Nobody Selling On-Page AI Optimisation Mentions
Muck Rack's "What Is AI Reading?" study has analysed more than 25 million links across ChatGPT, Claude, and Gemini in 17 sectors, over three editions between July 2025 and May 2026. Earned third-party media accounted for roughly 84% of AI citations, with the figure holding between 82% and 89% across editions. Your own site is competing for the remaining sliver.
Two independent analyses point the same way. A Search Engine Land study of about 25,000 of the most-cited URLs, drawn from nearly 400 million citations, found 63% pointing to ranked listicles rather than brand homepages. And research by Omniscient Digital covering 23,387 citations across five AI surfaces found reviews and social proof taking the largest single share, with brands' own product pages a much smaller slice. Three different methodologies, one direction.
The mechanism is not mysterious once stated. Someone asking an assistant which option is best is asking a comparison question, and a vendor cannot credibly answer a comparison question about itself. The system does what a careful person does: it reaches for a source with no stake in the answer. Optimising your own pages harder does not change the shape of the question being asked.
You have accepted that the majority of this work happens on domains you do not control, and budgeted accordingly.
Map the Sources That Already Get Cited
Do not start from a media list. Start from the assistants themselves: ask the questions your buyers ask, in clean sessions, several times each, and log every source that appears in the answers. The pages that recur across multiple assistants are your target list, selected by evidence rather than by a vendor metric.
- Reuse the prompt set you already have.
If you are tracking share of model, the citation column in that log is this list already. If not, build the prompt set the same way: fifteen to twenty-five buyer questions, never containing your brand name.
- Log the URL, not the domain.
Assistants cite specific pages. Knowing that a publication gets cited is not actionable; knowing that one particular roundup on that publication gets cited eleven times out of twenty is a brief with a target attached.
- Count recurrence across assistants.
A page cited by three different systems is drawing on something durable about its authority. A page cited once by one assistant may be a retrieval accident. Sort your list by how many distinct systems reached for it.
- Note whether you are already on the page.
You will often find yourself present but ranked eighth, or described with outdated information. Fixing an existing inaccurate entry is faster and higher-yield than earning a new inclusion, and almost nobody checks.
Expect the resulting list to be unglamorous. It is usually a handful of category roundups, two or three review platforms, a couple of directories, and a forum thread. That is what the citation studies describe, and it is a far more tractable target list than "build authority".
Qualifying the Targets Before You Spend Effort
Not every cited page is worth pursuing, and some are worth actively avoiding. The filter is whether the page reads like genuine evaluation: named tools, real trade-offs, criteria stated, no sign that every entry paid to be there.
| Signal on the page | Read | Action |
|---|---|---|
| Named criteria, honest trade-offs, some critical notes | Real editorial evaluation | High priority, pursue properly |
| Every entry positive, no criteria, affiliate links | Monetised list | Low priority; inclusion is cheap and worth little |
| Updated within months, dated entries | Maintained | Worth a correction request even if already listed |
| Last updated two years ago | Abandoned | Citation may persist, but a correction is unlikely to land |
| Openly sells placement slots | Marketplace | Avoid, see below |
Position within the list matters as well as presence. Assistants summarising a ranked list tend to carry the top entries forward more reliably than the tail, so being ninth of twelve on a cited page is worth considerably less than being third. That makes improving an existing placement a legitimate goal in its own right rather than a consolation prize.
Earning the Inclusion
The pitch that works is the one that makes the writer's page better. Editors maintaining a category roundup have a standing problem: their entries go stale, they lack detail on niches they do not use, and readers complain about omissions. Solve that and inclusion follows. Ask for a favour and it does not.
- Lead with a correction, not a request.
If the page describes a competitor's pricing wrongly or lists a product that shut down, say so, with sources. You have now been useful before asking for anything, and you have demonstrated you actually read the page.
- Offer the specific gap you fill.
Not "we should be on this list" but "your list has nothing for teams under ten people, which is the question in your comments three times". Editors add entries that close a gap; they ignore entries that duplicate one.
- Supply the entry pre-written and checkable.
A short factual description, current pricing, who it is not for, and links to substantiate each claim. Writers are working against time; the entry that requires no research gets added. Include the unflattering fit limits, because a writer who discovers you omitted them will remove you.
- Give them something to cite.
A number nobody else has is the strongest possible reason for a writer to mention you, and it works on assistants for the same reason. This is the argument for publishing original research even at small scale.
- Make sure the page you send them to holds up.
A writer who follows your link to a page with no pricing, no proof, and no clear description will quietly not include you. The page-side of this is the same work that converts arriving visitors.
The Review-Platform Half
Review platforms show up disproportionately in citation studies because they aggregate exactly what an assistant needs: many independent accounts of the same product, with structure. A complete, current profile with genuine reviews is table stakes, and a large share of small businesses do not have one.
The work here is unromantic. Claim the profiles that exist in your category, fill every field rather than the required ones, keep pricing and feature descriptions current, and build a habit of asking satisfied customers to leave a review. Detail matters more than the star average: a review describing a specific situation and outcome supplies matchable evidence, while "great product" contributes nothing an assistant can use to answer a constrained question.
Two rules keep this legitimate. Never incentivise a positive review specifically, which most platforms prohibit and which is detectable in aggregate. And never write reviews of yourself or arrange for others to, which is the fastest way to lose a profile permanently. For the local-business version of the same work, including how to ask without breaching platform rules, see how to get more Google reviews and local business AI search.
What Not to Buy
A market has appeared selling guaranteed placements on "AI-cited" listicles. It is the paid-link directory business with a new label, and it fails the same way: pages that sell inclusion stop being the disinterested sources that made them citable, and the marketplaces themselves create obvious patterns.
Think about what you would actually be buying. The value of a cited roundup comes from the assumption that its judgement is independent. A page where every entry paid has no judgement to be independent about, and once that becomes visible, whatever caused assistants to favour it erodes. You would be buying a position on an asset that your purchase helps devalue.
There is a policy dimension too. Paid links that pass ranking signals without disclosure fall under Google's link spam policies, and sponsored placements require disclosure that a careful editorial source will apply. Sponsorship is not the problem; pretending it is editorial judgement is. The recovery cost if this goes wrong is covered in our guide to link spam and disavowal.
Measuring Whether It Worked
Three numbers, monthly: how many of your target pages now include you, your average position on the ones that do, and your mention rate in the assistant prompt set. The first two are counts you control. The third is the outcome, and it moves slowly.
Do not expect a clean causal line between an inclusion and a change in citation behaviour. Retrieval is not a lookup table, models update on their own schedule, and an inclusion may take months to propagate or may never visibly move anything. Track the inputs because they are real and controllable, track the outcome because it is what you actually want, and resist the urge to attribute one to the other on a single month's data.
Pair this with the impressions view now available in the Search Console generative AI report and with referral attribution set up properly per fixing AI traffic in GA4. Between the three you get visibility, citation, and arrival, which is as complete a picture as currently exists.
Questions People Ask About Earning AI Citations
- Where do AI citations actually come from?
Overwhelmingly from sites the brand does not own. Muck Rack's ongoing 'What Is AI Reading?' study, which has analysed more than 25 million links across ChatGPT, Claude, and Gemini in 17 sectors, found earned third-party media accounting for roughly 84% of AI citations, with the figure between 82% and 89% across three editions from July 2025 to May 2026.
- Why do assistants cite listicles instead of brand websites?
Because a listicle answers the question that was asked. Someone asking which tool is best wants a comparison, and a vendor page cannot credibly provide one about itself. A Search Engine Land analysis of roughly 25,000 of the most-cited URLs across nearly 400 million citations found 63% pointing to ranked listicles rather than brand homepages, which is what you would expect from a system trying to answer a comparison question.
- How do I find which sources AI cites in my category?
Ask the assistants directly. Run the category questions your buyers would ask, in clean sessions, several times each, and log every source cited in the answers. The pages that recur across multiple assistants are your outreach list, and they are ranked by evidence rather than by domain authority scores.
- Can I pay to be included in AI-cited listicles?
You can, and it is the same trade that paid link directories offered a decade ago, with the same ending. Marketplaces openly selling placements create obvious footprints, and content that exists to sell inclusion tends to lose the credibility that made it citable in the first place. Sponsored inclusion also has to be disclosed, and disclosure is part of what a careful source is judged on.
- Is this just digital PR with a new name?
Largely yes, and that is a useful thing to notice rather than a criticism. What changed is the target selection and the measurement: instead of chasing publications by audience size, you chase the specific pages that assistants demonstrably quote in your category, and you measure whether your presence in generated answers changed. The outreach craft underneath is old.
Primary Sources
- Muck Rack: What Is AI Reading? study of 25 million-plus cited links across ChatGPT, Claude, and Gemini
- Omniscient Digital: Which content types LLMs cite most, 23,387 citations analysed
- Google Search Central: Spam policies, including link spam
- Google Search Central: Qualify your outbound links to Google
- FTC: Endorsement guides and disclosure requirements
