How to Get Cited by AI: A Practical Guide to Earning Brand Mentions

Short answer: Getting cited by AI means being named inside an assistant's written answer rather than ranked as a link beneath it. The citation data points at three levers: publish ranked comparison pages, because 63% of nearly 400 million AI citations point to listicles; get described the same way on other companies' pages, because those account for roughly a quarter of the citation pool; and confirm the retrieval crawlers can reach you, because a blocked crawler makes everything else pointless.

This is the tactical half of the story. For the strategic case — why clicks are falling while brand demand is not — start with AI search visibility. For the definitions, see what answer engine optimization actually is. This page is about the work.

Key takeaways

• Ranked lists are the format that gets quoted. Evertune reviewed the 6,000 most-cited URLs per model across six AI systems — about 25,000 unique URLs — and found 63% of nearly 400 million citations pointed to listicles. Between 71% and 86% of those were ranked lists, not unordered tips posts.
• Other companies' content is a bigger target than journalism is easy. Muck Rack classifies 84% of AI citations as earned media. Inside that number, journalism is 27% and corporate blogs and content are 24%. Your own site sits separately at 13.7%.
• The assistants do not behave alike. ChatGPT includes citations in 96% of its responses, Gemini in 82%, Claude in 55%. Being absent from one tool can have nothing to do with your site.
• Press releases work on trend questions, not buying questions. Roughly 1% of responses to industry-trend questions contain a press release citation — about 3.5 times the rate seen on best-of questions.
• LinkedIn Articles get cited. LinkedIn posts do not. Articles drive 46% of LinkedIn citations; regular posts drive 1%.
• Freshness is a filter. More than half of journalism citations come from the past year, so an un-refreshed page quietly ages out of the pool.

Can AI systems actually read your site right now?

Check this before anything else, because a blocked retrieval crawler makes every other tactic on this page worthless. Retrieval crawlers fetch a page so it can be quoted in a live answer. They are a different decision from the training crawlers that most block lists were written to stop, and plenty of sites have blocked both without meaning to.

The distinction matters more than it sounds:

The last two rows are the point. You can decline to become training data and remain fully quotable in live answers. Many sites made the opposite trade by accident, usually through a security plugin or CDN rule that treated every unfamiliar user agent as hostile.

This takes about a minute to confirm. Our AI crawler checker reads your robots.txt and reports which of these are permitted. If any retrieval crawler is disallowed, stop reading and fix that first.

Which pages get cited, and why is it almost always a ranked list?

Because a ranked list has already performed the comparison the model is being asked to perform. Someone asking an assistant which tool to buy is asking a comparison question, and the cheapest way for a model to answer it well is to lean on a source that already did the comparing.

The scale of this preference is larger than most content plans assume. Two independent studies, different methods, same direction:

The two headline numbers differ because they count different things: Evertune weighted by citation volume against the most-cited URLs, Wix Studio measured format share across a broad answer sample. The agreement that matters is directional. One format outperforms every other, and it is the one most content calendars treat as filler.

Two details inside the Evertune data are worth more than the headline.

The advantage is not uniform across models. Listicles made up between 40% and 65% of the most-cited URLs depending on the system, with Copilot at the low end and Gemini at the high end. If your buyers concentrate in one assistant, the format's value to you sits somewhere inside that spread rather than at the average.

Ranked beats unranked decisively. Between 71% and 86% of the cited listicles were ranked — "Top 5 CRM Tools" rather than "7 Ways to Save on Groceries." Numbering the list and committing to an order is doing real work, not decoration. A model answering "what's the best X" can lift a ranked entry and keep its meaning; an unordered tips post gives it nothing to rank with.

So the three page types worth building are narrow:

1. The ranked category list. "Best invoicing software for freelancers," numbered, with the comparison table high on the page rather than buried.
2. The head-to-head comparison. Two named products, feature by feature, with real prices and a stated verdict.
3. The alternatives page. Written for someone who already owns a competing tool and is unhappy with it — the highest-intent reader in the category.

You can write all three about your own category and include yourself. Everyone already occupying those answers is doing exactly that. The constraints are that you name where you lose, include the genuine contenders rather than a strawman field, use real figures instead of "affordable," and date your research. Paid and advertorial content accounts for 0.3% of AI citations, which is a fair proxy for how little weight overtly promotional material carries.

Does publishing on your own domain still work?

Yes, but less than you would like and differently than the headline statistic suggests. This is where the popular version of the advice gets sloppy, and the correction is the most useful thing on this page.

Muck Rack's study of more than 25 million cited links across ChatGPT, Claude and Gemini in 17 industries found that 84% of citations come from earned media. That figure gets quoted as proof that your own site does not matter. The breakdown says something more precise:

Read that carefully. "Earned media" in this study is an umbrella that includes corporate blogs and content — other companies' pages — at 24%, nearly matching journalism's 27%. Your own domain is counted separately at 13.7%.

Three conclusions follow, and none of them is "stop publishing."

Your own site is roughly one citation in seven. That is a real ceiling, but it is not zero, and it is the only part of the pool you control completely. It is also where the ranked lists from the previous section live.

Other companies' content is the second-largest earned category and the easiest of the big ones to enter. Getting into a journalist's story requires a news hook and a reporter's interest. Getting into another company's roundup, integration directory, or comparison page requires an email. Those pages are 24% of the pool, and most of them accept updates from vendors who make it easy.

Paid placement is the worst-performing option available. At 0.3%, buying an advertorial is close to buying nothing, in a channel where an unpaid mention on a mid-sized blog compounds indefinitely.

Where does the independent coverage come from?

Journalism is 27% of cited links, spread across more than 20,000 distinct outlets — so the target is wide rather than a handful of famous mastheads. That breadth is the opening for a company nobody has heard of. You are not trying to get into a national paper; you are trying to get into any of twenty thousand publications that cover your category.

One qualifier changes how you spend the effort: more than half of journalism citations come from the past year. Coverage decays. A feature from three years ago is doing far less for you now than it did then, which makes a steady trickle of small mentions more valuable than one big hit you never follow.

The free version of this work takes an afternoon:

1. Ask ChatGPT, Claude, Perplexity and Google's AI Mode your main category question — "what are the best tools for [what you do]?"
2. Record every source each one cites. Those are the pages the models currently trust in your category, and the list is usually shorter than expected.
3. Open each one. Are you listed?
4. Where you are not, email the author with something specific: what you do that nothing else on their list does, an offer of free access, and an honest note on where you are weaker than a competitor they already included. Most will ignore you. Some will not, and each one that adds you becomes a durable source that does not stop working when you stop paying.

Treat unlinked mentions as wins too. A sentence describing your company gives a model a usable association whether or not it carries a hyperlink.

Do press releases earn AI citations?

They do, but on a narrower question type than most wire-service pitches imply. Roughly 1% of responses to industry-trend questions contain a press release citation — about 3.5 times the rate on best-of questions. That is the whole shape of the opportunity: releases surface when someone asks what is happening in a market, not when they ask what to buy.

Which means the honest use case is narrow. If you want to be named in "best project management tools for agencies," a press release is close to the wrong instrument. If you want to be named in "what's changing in agency project management this year," it is a reasonable one — especially built around original data nobody else has.

Where you distribute also matters more than the pitch decks suggest. Among cited press releases, the newswires are not evenly represented:

That is a lopsided distribution to know before you buy. It does not mean the others are worthless, but it does mean the default choice has a measurable consequence.

The mechanics that make a release usable are unglamorous. Announce one specific, verifiable thing — a launch, a dataset, a partnership, a milestone with a number attached. Answer who, what, and why anyone should care in the first 50 words, and delete "we're excited to announce" before writing it. Spend your two or three anchored link slots on the pages you want quoted, not your homepage.

Which profiles and platforms are worth the effort?

Only a few, and the split within each platform is sharper than the split between platforms. This is the part most visibility advice gets wrong by recommending "be active everywhere."

LinkedIn: write Articles, not posts. Articles account for 46% of LinkedIn citations. Regular posts account for 1%. The daily posting habit that much of the industry treats as table stakes is contributing almost nothing to AI citation, while the long-form section nearly everyone ignores is doing nearly half the work.

Wikipedia: the citable pages are conceptual, not corporate. Cited Wikipedia pages skew heavily toward processes, technologies and methods (42%) and laws, regulations and standards (12%). Brand and company pages are 9%. If you have been treating a company Wikipedia page as a visibility goal, the data suggests the category page for the thing you do is the more valuable neighbourhood to be described in.

Reddit: a Gemini concern, not a universal one. Gemini draws about 2.4% of all its citations from Reddit; ChatGPT and Claude do not meaningfully cite it. Whether community presence is worth your time depends on which assistant your buyers use.

Google Maps: the overlooked one, if you are local. ChatGPT uses Google Maps at a rate of 188 citations per 1,000 queries. For any business with a service area, a complete and accurate Maps listing is a higher-yield hour than most content work, and it is almost never mentioned in AI SEO advice.

Across all of them, one rule holds: use the same description of yourself everywhere. A model reproduces the version of you that recurs most consistently across sources. Four adjacent-but-different bios give it nothing confident to say, so it names a competitor with a clearer story. Adjust length and tone per platform; keep the core phrase identical.

Why does your brand appear in one assistant but not another?

Partly because the assistants cite at very different rates. ChatGPT includes citations in 96% of its responses. Gemini does so in 82%. Claude does so in 55%.

That single spread explains a lot of confused reporting. A brand that appears steadily in ChatGPT and rarely in Claude may have no problem at all — Claude simply attaches sources to barely half of what it says. Before concluding that your visibility collapsed in one tool, check whether that tool cites anything in the first place.

Outputs also vary between runs of the same prompt, and citation behaviour shifts with every model update. A single query on a single day tells you close to nothing. Fix a set of ten buyer questions, run them monthly, record the date and model version alongside each result, and read the trend rather than any one session. The full measurement framework — what to track and what to ignore — is in our guide to AI search KPIs.

How do you write a page so a model can lift a passage out of it?

Write in self-contained units, because extraction takes one passage and drops it into an answer with none of the surrounding context. A paragraph that only makes sense in sequence cannot be extracted, and a page that cannot be extracted will not be cited however good the thinking behind it is.

Five changes, in order of how much they cost you:

1. Make headings the questions people actually ask. "How much does X cost?" rather than "Pricing." That phrasing is what the passage gets matched against.
2. Answer in the first two sentences under each heading. State the answer, then explain it. Never build toward it — a retrieved passage has about two sentences to prove useful.
3. Replace vague words with specific ones. "Reduces editing time from three hours to twenty minutes" survives extraction. "Saves time" does not.
4. Use absolute dates and show your sourcing. A passage stripped of context cannot resolve "last month." "Prices verified in August 2026" can travel.
5. Write table rows that stand alone. A model lifts a single row without its headers. A cell reading "Great for teams" is dead weight; "Teams of 5 to 20 publishing across 8 platforms" can be quoted directly. One well-built table can generate a dozen separate mentions.

The extraction test: take any single paragraph out of the page, hand it to someone with no other context, and ask what question it answers. If they cannot tell you, rewrite it.

Two structural jobs sit behind this. Add Organization schema on your homepage and FAQPage schema wherever you have questions — not because JSON-LD lifts citations directly, since the evidence for that is thin, but because it removes ambiguity about which company you are. Our meta tags checker will show you what is currently there. Then re-verify every price, claim and named competitor on your best pages once a quarter. Given that more than half of journalism citations come from the past year, the same recency preference almost certainly applies elsewhere, and refreshing a page that already carries trust is the highest-return recurring task available.

What should you do first?

Ordered by return on the hour, not by difficulty:

An honest expectation to set before you start. How long this takes depends on how contested your category is and how much presence you already had. Refreshing a page that already carries trust can surface in answers within weeks; building recognition from zero takes considerably longer. Anyone promising results next week is selling something.

One further caveat worth holding onto: the widely cited Pew Research finding on click behaviour — traditional results clicked in 8% of visits with an AI summary present versus 15% without, across 68,879 searches from 900 U.S. adults — was collected in March 2025. It is the best public dataset of its kind, and it is also getting old, gathered before much of the current AI Mode rollout. Directionally it has held up. Treat the precise figures as a floor rather than a current reading.

Frequently asked questions

What does it mean to be cited by AI?

Being cited by AI means an assistant names your brand, or links to one of your pages, inside the answer it writes for a user. It differs from ranking because there is no results page to place on — you are either among the two or three options the model puts forward, or you are absent from that conversation entirely.

What kind of page is most likely to get cited?

Ranked lists, by a wide margin. Evertune found 63% of nearly 400 million AI citations pointed to listicles, and between 71% and 86% of those were ranked rather than unordered. Head-to-head comparisons and alternatives pages follow. Company history pages and general tips posts rarely enter buying answers at all.

Do I still need to publish on my own website?

Yes, though it is one part of a larger pool. Muck Rack's breakdown puts owned media at 13.7% of AI citations, against 27% for journalism and 24% for other companies' blogs and content. Your own site is the smaller share, but it is the only one you control outright and it is where your comparison pages live.

Why does ChatGPT mention my brand but Claude does not?

Partly because the systems cite at different rates: ChatGPT includes citations in 96% of responses, Gemini in 82%, and Claude in only 55%. Retrieval sources and index freshness also differ. Track several assistants across several months before concluding anything from a single tool.

Should I block AI crawlers from my site?

Blocking retrieval crawlers such as OAI-SearchBot, Claude-SearchBot, PerplexityBot and Googlebot removes your ability to be cited in live answers. Training crawlers such as GPTBot and Google-Extended are a separate decision. You can decline to be training data while remaining fully quotable.

Are press releases worth paying for?

For trend coverage, sometimes. Around 1% of responses to industry-trend questions cite a press release, roughly 3.5 times the rate on best-of questions, so releases rarely help you get named in a buying answer. Among cited releases, GlobeNewswire accounts for 61%, PR Newswire 27% and BusinessWire 12%.

Does posting on LinkedIn help with AI citations?

Articles do; posts effectively do not. LinkedIn Articles account for 46% of citations from the platform, while regular posts account for 1%. If you are investing in LinkedIn for visibility in AI answers, that is the section to write in.

Does schema markup get you cited?

Schema clarifies which entity you are, which is worth the twenty minutes it takes, but the evidence for a direct citation lift from JSON-LD alone is weak. Treat Organization and FAQPage schema as disambiguation and hygiene, not as a lever that will move your mention rate by itself.

How long before any of this shows up in answers?

Months rather than weeks, depending on how contested your category is. Refreshing a page that already carries authority can appear in answers within weeks; earning recognition from a standing start takes considerably longer. More than half of journalism citations come from the past year, so ongoing work matters more than any single push.

Where to start

Run the one-minute version before you plan anything: open an assistant and ask who the best provider is for whatever you do. If you are not in the answer, you have your goal for the next 90 days.

Then check the mechanical precondition. Our free AI crawler checker tells you in seconds whether the retrieval crawlers can reach your pages, and the meta tags checker shows what a model sees when it does. If you would rather see the whole picture at once, the sample report walks through a full AI visibility audit on a real site.

Sources

1. Evertune via Search Engine Land — AI search loves listicles: what 25,000 URLs reveal about citations (May 2026)
2. Wix Studio AI Search Lab via Search Engine Land — AI citations favor listicles, articles, product pages (March 2026)
3. Muck Rack — Earned media still drives 84% of AI citations (May 2026)
4. Muck Rack — How AI citations have changed in the last six months (2026)
5. Muck Rack — 3 more key takeaways from our "What is AI Reading?" research (2026)
6. Muck Rack — FAQs on the May 2026 update of "What Is AI Reading?" (2026)
7. Pew Research Center — Google users are less likely to click on links when an AI summary appears (July 2025)

Every statistic on this page was verified against its primary source in August 2026.