LinkedIn
Most agencies won’t tell you this: the two loudest camps in AI visibility give opposite advice, and both can point to wins. Should you fix your own pages to earn citations, or get other people talking about your business? Smart people disagree because they’re watching different engines answer different prompts. An informational Google AI Overview needs a passage it can use. A ChatGPT or Perplexity recommendation needs reasons to name one business over another.
I used to tell clients a tight page rewrite was enough. I was wrong often enough to stop saying it. But the people who dismiss page work altogether are wrong, too. I’ll argue each side at full strength before ruling. My answer is not a split-the-difference budget: for Google AI Overviews, fix your pages first; for ChatGPT and Perplexity “best X” prompts, your own site is rarely enough.
The on-site case: give Google a passage it can lift
The on-site camp starts with retrieval. Google’s AI Overviews draw on pages in Google’s index, so a page that ranks for the main query and its related follow-ups has a path into the answer that an invisible page does not. A Surfer analysis of 10,000 keywords found a 0.77 correlation between fan-out coverage and citation likelihood. Pages covering the main query plus at least one fan-out drove 51% of citations, against about 20% for pages covering only the main query.
That’s a strong argument for covering the question people ask and the follow-up they are likely to need. It isn’t proof that adding one paragraph causes a citation. It does tell me where to look before I commission another roundup: does the page answer enough of the conversation to be useful?
Then comes extraction. Ranking gets the page into the room. A clear, self-contained passage gives the engine something to cite. That is the logic behind a direct 40-to-60-word answer under a matching H2 or H3, question-based headings, useful tables and substantive FAQs. The answer should stand on its own. If a reader needs seven paragraphs of setup to understand it, the engine has the same problem. Practitioners document this answer-first pattern for AI Overviews because a block that needs no added context is easier to reuse than a conclusion buried in a winding service page.
The technical floor is less glamorous. A page must be indexed and eligible for a snippet; there is no separate AI Overview markup requirement. That does not make crawlability optional. If the bot cannot load the page or reach the answer, immaculate prose in a content document will not help. I have spent enough time watching good landing pages fail because the live version differed from the tidy version everyone approved. Check the page that exists, not the one in the brief.
This camp’s best prompt is informational: how to reset a shutoff valve, what a term means, why a process works. The engine can answer by finding a relevant, accessible passage rather than comparing claims about competing businesses. Listicles account for about 27% of cited content in the observed data cited by this camp; their headings and compact explanations make them easy to scan and lift. That does not mean every answer should become a listicle. It means structure earns its keep when the question has a direct answer.
If your page already ranks but never appears for a definitional prompt, don’t start by chasing mentions. Read the page under the matching heading. Can you find the answer in its first two sentences? One prompt, one accessible page, one answer block worth quoting. That is the on-site case at its strongest.
The off-site case: a recommendation needs more than your word
The off-site camp begins where the on-site case gets uncomfortable: comparative prompts. When someone asks for the best bookkeeping tool for a 10-person agency, the engine is not merely looking for a tidy definition. It has to choose among businesses making interested claims about themselves. Every vendor can write “best for agencies” on its own site. That is not much of a recommendation.
Reviews, forums, listicles, press and videos give the engine other accounts to consider. If several sources independently describe the same fit, a recommendation has more to draw on than a company’s own copy. Your site explains your offer; other people help establish whether the claim holds up. That distinction matters most when the prompt asks for best, top or versus, rather than what is or how to.
The association between mentions and visibility is hard to ignore. An Ahrefs study of roughly 75,000 brands found branded web mentions correlated at 0.664 with AI visibility, compared with 0.218 for backlinks. Brands in the top quartile of mentions averaged 169 AI Overview appearances against 14 for the next quartile. The same analysis put YouTube mentions at about 0.737. Those figures do not promise that collecting mentions will produce a particular number of citations. They do show why a page-only strategy leaves something substantial out.
Reddit makes the point in less abstract terms. Its threads provide people’s comparisons in language closer to the prompt than most product pages use. One cross-engine analysis counted 23.6 million cited Reddit pages across AI answers. You can dislike a thread’s verdict. You cannot make it disappear by polishing your H2s.
There is also a useful distinction inside the off-site argument: a listing is not the same thing as an independent review. A business can control its listing details; it cannot turn those details into someone else’s endorsement. Both still give an engine places beyond the business’s website to find and check information. A Yext analysis of 6.8 million citations found 86% came from brand-controlled sources, split between first-party sites at 44% and listings at 42%. OpenAI leaned more heavily on listings at 48.7%, while Gemini favored websites at 52.1%.
That finding complicates the slogan “AI trusts independent consensus.” Good. Slogans deserve to be complicated. The stronger off-site case is that, for a recommendation, your website is only one surface in a much wider account of who you are, what you do and why someone might pick you. Listings make you findable outside your domain. Detailed reviews and discussions supply the judgment your listing cannot write for itself. A beautiful page with no echo elsewhere still has to argue its own case.
Where each case breaks down
The on-site camp cannot explain away the great page nobody quotes. Picture a SaaS page ranking near the top for a how-to query: fast load, a clean answer under the right heading, no obvious extraction problem. Google’s AI Overview still cites two other pages. For the “best tool for X” version of the prompt, ChatGPT does not mention the business at all. A clean page improves its chances; it does not oblige an engine to choose it, and it does not provide independent reasons to recommend the product. The Ahrefs analysis found 26% of tracked brands had zero AI visibility despite having websites. Having a site is not the same as being part of the answer.
The off-site camp has the mirror embarrassment: the brand people discuss that never gets linked as the source. Reviews name it. A roundup includes it. Then an informational AI Overview cites a competitor’s duller page because that page actually answers the question. The talked-about brand may have buried the definition, put it in a video without a transcript, or made the relevant page difficult to load. Other people can speak well of you without making your own page useful to a crawler.
I used to think buzz would pull a weak page along. That is too generous to the page. Mentions cannot supply a passage that your site never makes available. Equally, a passage cannot supply the outside judgment a recommendation calls for. Neither camp has a universal fix; each has a prompt where its mechanism matters most. That is the consideration that decides my verdict.
My verdict: fund the bottleneck for the prompt you’re losing
For Google AI Overviews on informational prompts, the on-site camp wins the first round. The engine needs a relevant page it can access and use. In the 10,000-keyword study, pages covering both the main query and at least one fan-out took more than half the citations. If you are missing how-to and definitional prompts, inspect your pages before you buy press. Does a relevant page appear in search? Can the bot load it? Is there a short answer under a heading that matches the question? Fix the failed step first.
For ChatGPT and Perplexity “best X” prompts, the off-site camp wins. Your website is an interested party. The engine can also draw on listings, reviews, videos, roundups and discussions, so a site rewrite alone is a thin response to a recommendation problem. The citation analysis showing OpenAI’s heavier reliance on listings does not make every listing a trusted review. It makes the practical point: if your information and reputation exist only on your own domain, you have left much of the answer to other businesses.
That is not a call to abandon your site. It is an order of operations. A recommendation can still send someone looking for a clear page about the business. An informational answer can still favor a brand the engine recognizes elsewhere. But when time and budget are limited, “work on everything” is not a decision. Find the missing citation, identify what kind of answer it requires, and fund the constraint that kept you out.
Informational prompts need a passage; comparative prompts need a case
The engine split becomes easier to use when you look at the prompt. An informational question asks for a fact, definition or step. A self-contained passage gives the engine an answer it can use with little explanation. A comparative question asks for judgment. It invites the engine to assemble an account from several sources, including ones you did not write.
If the missing prompt starts with how, what or why, inspect the page first. If it starts with best, top or versus, inspect the echo. That is a sorting rule, not a claim that every engine answers every prompt identically. Keep the actual missing citation in view.
Think of it the way I think about PPC. Landing page quality and bid strength both matter, but only one may be stopping you right now. A great Quality Score with a $2 bid can miss the auction. A $50 bid with a broken landing page can buy expensive bounces. For an informational Overview, the page is often the first constraint to check. For a ChatGPT best-list, start with what sources beyond your site say about you. Eventually you may need both. You fund the bottleneck first.
What to do Monday: sort the losses before you spend
Pull the last 10 prompts where a competitor appeared and you did not. Write each on its own line. Mark I for informational or C for comparative, and note which engine produced the answer. Do not combine a Google how-to Overview and a Perplexity “best X” answer into one visibility score. They are different losses, with different first places to look.
Then give each line a 15-minute check:
- For an I prompt, inspect your page. Search the query in Google. Does a relevant page rank in the top 10? Can the bot load it? Is there a 40-to-60-word answer under a matching heading? Note the first point of failure rather than rewriting everything at once.
- For a C prompt, inspect the outside account. Run the prompt in ChatGPT and Perplexity. Which independent sources name the winner, and which name you? Check your listings separately from reviews or discussions; complete business details are useful, but they are not independent praise.
- Choose the larger actionable backlog. If six losses are I prompts and four are C prompts, the six point to page work and the four to off-site work. That does not guarantee the six will be easier wins. It does stop you paying for the wrong kind of fix. Ten checks at 15 minutes each take more than an hour; the initial sorting is quick, the diagnosis takes work.
Say you spend $20k a month on search. That is a good reason to do the sorting before funding another content or reputation project, not evidence that one project will earn a particular return. Record the prompt, engine, current citation and planned change so the next check tells you whether anything moved. Otherwise “we improved AI visibility” can mean someone saw your name once in a screenshot.
- Fix the I backlog on your site: a relevant page for each missing question, a concise answer block, tables where they clarify a comparison, and a dated retest in Search Console and an AI check.
- Fix the C backlog off your site: complete listings with consistent names and categories, requests for detailed reviews that mention the job done, and opportunities for specific third-party coverage or discussion.
- Do not buy ten directory links and call it consensus. A thin citation may repeat your name without helping anyone explain why you fit the prompt. Look for sources that actually say something useful.
That exercise is where teams tend to stall. Page rewrites sit in a document. Listing details drift. Review requests stay on someone’s to-do list. Monitoring the answer is useful, but watching a competitor get cited again is not the same as changing the conditions that put them there. That is why I point people toward tools that help you get featured, not just tools that count mentions.
groas takes the execution burden of paid and organic search off the client’s dashboard, with autonomous work running continuously and a named human strategist responsible for direction and guardrails. That model is a better fit for this problem than paying for another periodic report and hoping someone acts on it. But the first decision is still yours: if your missing list is mostly informational prompts, start with your pages. If it is mostly best and versus prompts, start with your echo. Fund the bottleneck until it stops being the reason you lose.