How to get cited in ChatGPT, Perplexity, and Gemini

Write the questions a buyer asks before they book or buy, not the questions a marketer wishes they asked. Pull them from lost deals, won deals, and the first three minutes of sales calls. A D2C brand or a Shopify store has the same raw material in support tickets, product reviews, and the questions shoppers ask in chat. You want four clusters. Category: "best tools for X." Alternatives: "what to use instead of Competitor." Fit: "best X for a team that already uses Y." Constraint: security, price band, geography, implementation time. Twenty to forty prompts is enough to run a program. Two hundred prompts is a way to avoid deciding.

Keep the wording clumsy if that is how buyers talk. Assistants are asked messy questions. A prompt polished into a blog headline measures a query nobody types into a chat. Include the competitor's name when the competitor is actually in the deal. Include the job title only when it changes the answer. A CFO and a staff engineer do not get the same shortlist, and you should not pretend they do if you only sell to one of them.

Run the set once before you change anything. Save the full answer, the order of names, every URL, and every factual mistake about you. This baseline is the asset. Later arguments about "I think we showed up" die when the baseline is a file. The GEO agent reruns the set every week on eight engines, asking each prompt several times so one odd answer is not mistaken for a trend, and the file stays current without a person pasting screenshots into a slide.

Group every cited URL by domain and by job. You will usually find a pattern within a single category. A few commercial comparison sites. A documentation page. A community thread where practitioners argue. A review marketplace. Sometimes a news piece that is already stale. Count how often each domain appears across the prompt set, and note which engine likes which domain. Perplexity may cite a hands-on review that ChatGPT never mentions. AI Overviews may cite a page that already ranks. The overlap is your short list. The uniques are engine-specific tasks. The GEO agent keeps this count for you, by domain, by source type, and by engine.

Open the pages. Do not score them by domain rating and walk away. Read whether they name you, whether the paragraph is fair, and whether the page is the kind of URL that will still exist in six months. A citation to a thin affiliate page is a fragile win. A citation to a careful comparison that gets your limits right is a description you can live with. Write down the sentence you wish the page had used. That sentence becomes the pitch to the editor, or the outline of the page you will publish yourself.

Ignore sources that cannot move a buyer even if an engine likes them. A student blog, a scraped directory, a page in the wrong language for your market: presence there can create a mention and still embarrass the brand. Geographic and language filters matter. If you sell in the UK, a US-only roundup that cites dollar pricing may be teaching the engine the wrong offer. Track prompts by country when the answer changes by market.

The page you control should be easy to lift a sentence from. Open with the answer: who the product is for, what it does, and the tradeoff. Use the buyer's noun in the heading. Put numbers, limits, and supported integrations in prose, not only in an image. Models read text. A pricing table rendered as a picture is invisible. A short paragraph that states the price band and who it fits is citable.

Write the comparison you wish a journalist would write, including the cases where you are the wrong choice. Engines reuse pages that sort the market. Pages that claim you win every row teach the model nothing, because the claim is not information. State the incumbent's real strength. State your real strength. Name the company size, the stack, and the job. Internal links should point to the integration page, the security page, and the use case, so a follow-up question has somewhere true to land.

Structured data supports this. Organization, Product, FAQ, and Breadcrumb markup make the page easier to parse. They do not persuade a model to recommend you. The GEO agent builds the markup straight from your approved facts, so the markup and the page say the same thing. Ship the markup. Do not confuse it with a strategy. The answer-first page guide goes through the shape of the page itself.

Pick a facts block and paste it into the places you control: homepage, about page, docs intro, profiles, app marketplace listings, partner one-pagers. The block is boring. Category. Buyer. Job to be done. Proof you can show. Limit you will admit. Last reviewed date. When those lines match, an engine stops hedging. When they conflict, the answer picks a side at random or describes you as "a platform that offers a range of solutions," which is how models talk when they are lost. In the GEO agent, this block is the facts file, kept with its history, and every answer is checked against it.

Then correct the places you half control. Claim profiles on review sites. Update the outdated listing. Reply where a community thread states a deprecated limitation as if it were current. A correction in the original thread is more useful than a new post that argues with a ghost. Keep a log of what you changed and the date, so the weekly retest can be read against your own edits.

Fintech, health, and other regulated categories should treat accuracy as the primary KPI. A flattering wrong sentence about coverage, licensing, or pricing is worse than silence. The GEO report should flag a wrong description even when the mention rate went up, and the GEO agent does: each issue keeps the exact quote, the engines that said it, and how severe it is. Sales will feel that error before marketing celebrates the chart.

When an engine states a wrong fact, the fastest fix is rarely a new page. It is the page the engine learned the claim from. Start with the citations on the answer itself, since engines often quote the exact passage. Then search the web and the news for the wording of the claim, and check reference sources such as Wikidata. The page that repeats the claim, shows up for the prompt, and is cited by more than one engine is usually the one to fix first.

Fix in order of control. Your own pages come first, because the edit is yours to make the same day: write the exact before-and-after text, not "update the pricing page." Profiles you control come next, such as review site listings, marketplaces, and directories. Publishers come last, with a short, polite correction request that names the page, quotes the wrong sentence, gives the right one, and links to the proof. One open request per domain is plenty, and a few per quarter is the most anyone should send.

Then check that the fix held. Re-ask the prompts that produced the claim on the engines that made it. One clean answer can be luck, so wait for two in a row before calling it fixed, and reopen the task if the claim comes back. The GEO agent runs this loop: it traces each issue, drafts every fix for your approval, and leaves the sending to you, from your own mailbox. Then it rechecks the engines until two clean checks verify the fix, and alerts you if the claim reappears.

Once you know the cited URLs, outreach has a target. Pitch the editor, the list author, or the reviewer with the correction or the missing row, not with a generic guest post. Show the prompt, the fact that their page is the one being quoted, and the specific sentence that is stale. People who maintain comparison pages care about being wrong in public. They rarely care about your brand narrative.

Earned coverage is the default because the sentence is written by someone the engine has already decided to trust. A reviewer who has used the product, a practitioner who implemented it, a newsletter that covers the category: those mentions travel. Digital PR that earns a real explanation will outperform a syndicated blurb that says nothing. The same mentions often help classic rankings, which is why this work should sit on the shared SEO and GEO list rather than in a separate "AI PR" silo.

Where earned coverage is slow and the cited publisher sells labeled sponsored slots, a paid placement can put your facts on a domain the engine already visits. Label it. Do not ask for a fake editorial disguise. Do not expect the invoice to guarantee a citation. Engines cite pages, and they drop pages. Judge the placement by the next month of prompt retests. If mention rate and accuracy did not move for the prompts that cited that domain, stop buying that domain. Log every placement, too. The GEO agent watches placement pages weekly and flags one that changes, moves, or disappears.

Perplexity and other live-search assistants tend to move first. They are built to fetch current pages and show them. A new comparison, a corrected review, or a docs page that finally answers the question can show up in the citations within weeks. That speed is useful and easy to overread. A single good week is not a moat. Keep the prompt set fixed so you can tell a real shift from a one-off retrieval.

Google AI Overviews lean on pages that already have a reason to be retrieved for that query. Strong SEO on the exact question, plus a page structure that answers it in the first screen, is the lever. If you do not rank and you have no cited third party, an Overview will not invent you out of kindness. Technical health, internal links, and a page that satisfies the query are the work. The technical SEO guide covers the foundation those overviews still depend on.

ChatGPT and Gemini can mix browsing with longer memory. When browsing is on, they behave more like a live search and can reflect new sources sooner. When the answer is coming from accumulated descriptions, you need repeated, consistent mentions across sources the model has had time to absorb. That is a quarter-scale job, not a launch-week job. Report the engines separately. A blended "AI visibility score" hides the only fact that changes the plan: which engine moved.

Rerun the same prompts on the same engines, more than once each. Diff the mentions, the position, the sentiment, and the cited URLs against the last few weeks, and only call it a move when it is bigger than the usual swing. Write three lines: what moved, what you shipped that could explain it, what you will ship next. Attach the leads that named an assistant as the source. If you cannot attach leads yet, attach branded search trend and the conversations your website and inbound agents tagged. Measurement without a commercial echo becomes a hobby.

Shift effort toward the engines and prompts that still lag and still match pipeline. A prompt you already win does not need a fifth article. A prompt tied to your largest deal type, where a competitor is the only name in the answer, deserves the next page and the next pitch. This is the opposite of a content calendar that fills Tuesdays. The calendar is the prompt gap list.

Keep a human on the description. Models will compliment you for qualities you did not claim, and they will repeat a competitor's attack if that attack is the only sentence on the web. Read the actual paragraph. If it is wrong, the task is a correction, not a celebration of inclusion. Inclusion with a false feature is a tax on the sales team.

Stop publishing paraphrases of your homepage across a network of empty blogs. Scaled near-duplicate pages are a poor citation strategy and a rankings risk. One specific page beats twenty vague ones. Stop buying expired domains and fake editorial links to "boost AI." The engines are looking for a page that answers the question, not for a trick that used to move a toolbar PageRank.

Stop changing the prompt every week to chase a flattering answer. A moving prompt set cannot show progress. Add a prompt when the pipeline grows a new question. Retire one when you exit a segment. Otherwise the list is frozen so the chart means something.

Stop treating citation as a channel you can fully own. You will share the answer with competitors. The goal is a fair sentence and a source you trust, repeated on the prompts that create revenue. When that is true, the next search, the next call, and the next demo have your name in them. The measurement guide shows how to report that without fooling yourself.