There is no magic prompt that gets you cited by ChatGPT or Perplexity. Stop collecting lists of them. You do not get cited because you guessed what the buyer typed. You get cited because your page answered something the engine searched for after they typed it.
ChatGPT rewrites a buyer question into one or more targeted queries for its search partners, then runs narrower follow-ups after it reads the first results. In one logged sample of 17,962 live searches, only 5.3% repeated the prompt word-for-word. I used to build match-type structures around exact wording in Google Ads. I was wrong to bring that habit to AI search. The wording is not the lever.
Perplexity works the same way in principle and differently in practice. It searches live on every question, pulls ranked results through its own crawl plus third-party indices, fetches full page content, then attaches numbered citations to claims. Typical density runs around nine to ten sources per answer versus about three for ChatGPT. That sounds generous until you realize what it means: Perplexity needs more pages to build an answer, so it reaches further down the trust list.
Ask both engines the same question and they cite almost different webs. Across 22.7 million citations in the first half of 2026, 89.1% of sites ChatGPT cited were never touched by Perplexity, and vice versa 90.1%. In one contractor test of 1,429 answers, Perplexity cited LinkedIn 246 times versus 6 for ChatGPT, while YouTube showed up 214 times in AI Overviews versus 4 for ChatGPT. These are different retrieval environments, not one results page with three coats of paint. One prompt list cannot cover them.
Free tool: the AI Visibility Checker asks ChatGPT, Gemini, Claude, Perplexity and Grok the question your customers ask and shows whether they recommend you.
That is why I tell clients to forget the prompt they want to rank for. The prompt is raw material. The hidden sub-queries are the contest. Learn how ChatGPT picks a source from the search you never see and you stop asking the wrong question.
A prompt list is not keyword research
Asking what prompts lead to citations in ChatGPT versus Perplexity feels like keyword research. It is not. Keywords had a search volume and a results page you could open. Prompts have neither. In a SparkToro test, 600 volunteers ran 12 prompts 2,961 times across ChatGPT, Claude and Google AI. The test found less than a 1-in-100 chance the same prompt returned the same brand list twice. The order was even less stable, around 1 in 1,000. Someone selling you the 50 prompts buyers use is selling a snapshot of something that may not repeat tomorrow.
Think of it the way I learned to think about broad match. One buyer sentence fans out into several engine searches you never see. OpenAI describes this as rewriting the prompt into targeted queries, and outside logs make the spread visible: 27% of searches added the year the user never typed, and 12% asked for a named vendor the user never named. Only that 5.3% sliver I mentioned earlier kept the wording intact. Match your page only to the buyer’s exact sentence and you ignore most of the searches in that sample where a citation could be decided.
A prompt list also hides the most useful distinction: the buyer’s question and the engine’s need for evidence are not the same thing. A buyer may ask for a recommendation. The engine may go looking for price, setup requirements and comparisons before it names anyone. Rewriting your page to echo the recommendation prompt does not answer those narrower questions.
Your buyer asks one question and the engine searches for the parts behind it. You cannot control what strangers type into a box. You can control whether your page is the clearest answer when the engine goes looking for proof of each sub-claim.
The citation is decided after the buyer hits enter
Say a buyer types, “best CRM for a 20-person plumbing company that hates data entry.” The engine can split that into smaller searches: which CRMs suit small home-service firms, which require the least manual entry, and what they cost at 20 seats. ChatGPT and Perplexity can take different routes through those questions. Perplexity fetches pages live; ChatGPT runs its own searches and follow-ups. The final answer stitches together what those searches retrieve, with citations pointing to pages that support its claims.
Your homepage optimized for the full prompt may never enter that contest. Even if it does, a homepage saying “built for growing service businesses” does less work than a page that answers what happens at 20 seats. The engine needs something it can use, not a slogan that sounds vaguely adjacent to the question.
Why the two engines name different winners comes down in part to where they shop. ChatGPT uses its search partners and its own index preferences. Perplexity uses its PerplexityBot crawl plus third-party indices and pulls more sources per answer. Reddit shows the split clearly: tracking published in October 2026 found ChatGPT Reddit citations fell 88% from March to September while Perplexity still cited Reddit in about 20% of answers. Same buyer question, different shelves, different citations. Stop writing for the prompt. Start winning the sub-queries behind it.
Three kinds of AI-search advice I’d stop buying
1. Prompt lists dressed up as keyword lists
I get why agencies sell them. A spreadsheet of 200 prompts feels like work product. But a prompt list tells you what a tool guessed buyers might type, not what the engine searched when a real buyer typed it. Frequency across many runs is more stable than any single run, but even that stability describes the test, not your buyer.
Use a buyer question to start an investigation, not to define the page you will write. Look at the cited pages, ask what narrower question each one answered, and note which questions keep coming back. Log the sub-queries if you can. Bin the prompt list. A longer list just gives you more wording to track while leaving the retrieval problem untouched.
2. “AI-friendly” formatting without an answer worth citing
You know the checklist: a 40-to-60-word direct answer under a question heading, FAQPage and Organization schema matching visible text, third-party mentions on review sites and roundups the engines trust. That advice is repeated in almost every AI-search guide. Clean structure helps retrieval. It does not earn a citation by itself.
I have seen tidy pages with perfect schema sit uncited for months because no sub-query needed what they said. The page was easy to parse and hard to use. If the buyer’s hidden question is about migration, a beautifully marked-up paragraph about “seamless workflows” still does not explain migration. Answer the specific question first; make the answer easy to extract second. Reversing that order produces immaculate pages with nothing to quote.
3. A visibility score treated as the goal
Tools run 25 to 300 synthetic prompts, average who got named, and call the average a score. Critics call these numbers vanity metrics for a reason: the score is not reproducible and does not say what to fix. That matches what I see in accounts. The tracker tells you whether you appeared in its test. It does not tell you which hidden search you lost or which paragraph would have won it.
A score can go up while the page remains useless for the buyer question that matters. It can go down because a fresh run returned a different brand list. Neither movement tells an editor what to change on Tuesday morning. Paying a retainer to move that number is paying for someone to watch the weather.
Your tracker sees the prompt, not the searches behind it. If your agency reports the score without showing the sub-queries and the page that failed each one, they are selling prompts. The work is elsewhere.
Replace prompt tracking with pages that answer the fan-out
Start by mapping buyer questions to the sub-queries behind them. Take five questions that keep coming up on sales calls. For a plumbing CRM, those might concern price at 20 seats, setup time, how dispatch works and what breaks during migration. Ask ChatGPT and Perplexity each question five times. Write down the sources they cite, not just the brands they name in prose.
Then sort those citations by the question each page appears to answer. Is a directory supplying pricing? Are vendor docs supplying setup steps? Does a third-party page keep appearing when the engine compares options? You are not trying to reconstruct a hidden search log from a citation. You are looking for repeated evidence needs that your own pages fail to meet. One page that answers a recurring sub-query beats one pillar page that merely repeats the original prompt.
I run this like an old search-term mining routine. In Google Ads I used to pull the search terms report every Monday at 1am and add the junk as negatives. Here I pull cited URLs and ask which narrower question each one won. If three answers cite a directory for pricing but vendor docs for setup steps, I know to inspect the pricing page separately from the how-it-works page. They are not losing the same contest. After repeated runs, the aim is to make an informed bet about the fan-out, not pretend you can dictate it.
Next, make one paragraph per page that survives being quoted. Engines cite a page to support a claim, but the useful material may be only a sentence or two. Each money page needs a self-contained block: the direct answer, the relevant detail and the condition that limits it. If your actual setup data says a 20-seat shop importing from ServiceTitan takes 11 days on average because dispatch rules must be rebuilt, say that plainly and show the condition. If you cannot support that number, do not put it on the page. “Streamlined onboarding” is not a substitute for an answer.
Read the paragraph by itself before you publish. Does “this” still have a clear referent? Is the price tied to the right number of seats? Can someone tell whether the setup description applies to every customer or only to an import from ServiceTitan? Fix those gaps in the paragraph, not in a footnote five screens away. If the answer breaks when you lift it out of context, it is a poor candidate for a citation.
Then get named on the third-party pages the engines already cite. This is the unglamorous part nobody wants to pay for. Perplexity leans hard on Reddit, LinkedIn, YouTube and roundups, while ChatGPT has separate favorites. You do not fix that with phrasing on your homepage. You need to be present and specific where the engine already shops: a pricing page a directory can quote without guessing, a spec sheet YouTube reviewers can read on camera, or a customer story with headcount and migration detail that a Reddit thread can link without embarrassment.
That work has a cost. You have to make details public, keep them accurate and maintain pages beyond your own site. A paragraph on your domain cannot stand in for a missing third-party mention. Nor can ten tweaks to a blog introduction. If a cited roundup is answering the comparison question and your brand is absent, polishing your FAQ schema will not put you in that roundup. Work on the source the engine uses for the claim you want to win.
Ask your agency for edits, not a visibility chart
Ask for four things from last month:
- The buyer questions they tested.
- The narrower questions that kept recurring in cited results.
- The exact paragraph or page they changed in response.
- The citations they saw afterward, including what did not change.
An agency doing the work can put that on one page. An agency selling prompts has a visibility chart and a fresh list of 200 prompts to track next month. I learned this distinction the expensive way running ads. Anyone can report CPA. The person working in the account can tell you which keyword, bid and landing page moved it. AI citations demand the same operational detail, even if the results are less stable from one answer to the next.
If there is no edited page and no before-and-after in actual answers, ask what the retainer bought. groas treats this as execution, not reporting: new pages and technical fixes every day under its SEO and AI Search service, aimed at the sources the engines already trust. Prompts are cheap. Edits that survive quoting are the product.
Skip the citation chase if the basics are missing
This will not work for everyone, and here is who should skip it for now. If you get fewer than, say, 20 qualified sales conversations a month from search, fix your offer pages and tracking first. If buyers cannot tell what you charge, how fast you onboard and who you are for, you do not have a prompt problem. You have pages that fail the ordinary questions a buyer asks before choosing anyone.
I would also skip any retainer that promises to get you recommended in ChatGPT answers without touching third-party listings. Keep your money. Build one quotable paragraph per money page, get listed where your category already gets cited, and retest in six weeks. That is less glamorous than a dashboard. It also leaves you with pages that answer something useful even when the next AI response changes.
Keep chasing the magic prompt and you get exactly what prompt lists deliver: a report that moves while pipeline does not. Someone types a new variation next week, the fan-out shifts, and your tracked wording is history. Meanwhile, a competitor with a plainer page answers the sub-query you ignored and takes the citation you were chasing.

