Write for queries, not topics: how AI search decides what to cite

Most advice on AI content strategy still sorts your pages into two bins: commodity content that anyone could write, and unique content worth publishing. That sorting is fine for deciding what not to write. It is the wrong test for deciding what AI will cite, because AI does not rank your page against other pages. It assembles an answer to one specific query, and then decides whether your page is worth pulling into that answer.

Here is the gap that creates. A polished page on a broad topic can be technically excellent and still get skipped, because the model already knows the generic answer and writes it directly. Ahrefs analyzed 300,000 keywords and found that when an AI Overview appears, the top-ranking page loses about 34.5% of its clicks on informational queries (Ryan Law and Xibeijia Guan, April 2025). The page is not losing to a better page. It is losing to the answer.

This post reframes the query you should be asking about every piece of content. Not “is this page unique,” but “is the current best answer to this specific query still generic.” That second test is easier to act on, it points you at queries you can actually win, and it maps to how engines like ChatGPT, Perplexity, Gemini, and Copilot decide what to cite. I run this reframe across audits at BlueJar, and it changes what gets written first.

AI answers queries, not topics

Think about what actually happens when someone asks an assistant a query. The model does not retrieve a topic. It retrieves an answer to the exact thing that was asked, often by pulling a few specific passages from a few specific pages.

That distinction matters because demand lives at the query level too. “Schema markup” is a topic. “Does schema markup change whether ChatGPT cites my page, and by how much” is a query. The topic has a thousand pages competing for it and a model confident enough to answer it without anyone. The query has almost no good published answer, so whoever writes it owns it.

Zero-click data shows the same thing from the user side. Pew Research Center tracked 68,000 real searches and found people clicked a result only 8% of the time when an AI summary appeared, compared to 15% without one (reported by Search Engine Journal, July 2025 study). The clicks that survive are the ones where the on-page answer is better than what the model could say on its own. Generic topics fail that test. Specific queries pass it.

This is also why a single visibility number hides the real story. If you measure at the page or domain level, you get one blended score. Measure at the query level and you can see which exact queries name you, which name a competitor, and which the model answers without citing anyone. That is the difference between knowing you have a problem and knowing which problem.

The commodity-versus-unique test sorts the wrong thing

The commodity-versus-unique frame asks: could someone else have written this page? It is a page-level query, and it leads to page-level answers like “add original data” or “cover more of the topic.” Useful, but it misses the actual mechanic.

The better frame asks a query-level query: is the current best answer to this specific query still generic? If the best result a model can assemble today is a vague, hedged, could-apply-to-anyone paragraph, that query is wide open. You do not have to out-publish a category leader. You have to be more specific and more useful than a generic answer, on one query, at a time when no one else has bothered.

Consider the contrast directly. The work, the moat, and the AI behavior all change as you move from topic to query.

Read the bottom row again. Broad pillar pages are the easiest content for AI to replace, because the broad answer is exactly what a model is built to summarize. The narrow, specific answer is the part it cannot make up, so that is the part it has to cite. The reframe is not a style choice. It tracks how the systems actually work, which is the same argument we make about GEO versus SEO: the inputs overlap, but the unit you measure and optimize changes.

Why specific answers get cited and broad pages get summarized

The mechanism is retrieval. Engines that ground answers in live results, like Perplexity, pull a handful of candidate pages per query and quote a few of them, scoring on relevance, freshness, and how cleanly an answer can be extracted from the page. A broad pillar page buries the answer to any single query inside ten other sections. A page built to answer one query puts the answer where it can be lifted in one pass.

There is a second reason, and it is the one most content strategies miss: engines disagree about sources. Averi reported that only 11% of domains cited by ChatGPT are also cited by Perplexity (2026). The same query, asked of two engines, pulls almost entirely different sources. A blended “AI content strategy” aimed at no query in particular has no way to act on that. A query-level approach does, because you can see which engine names you for which query and fix the gap per engine. That is exactly why BlueJar runs the four engines (ChatGPT, Perplexity, Gemini, Copilot) separately rather than reporting one score. The fix is engine-specific too, which is why getting cited by ChatGPT is a different job from getting pulled into Perplexity.

Two failure modes only show up at the query level, and they need opposite fixes:

• Cited but not mentioned. Your domain sits in the source block, but your brand never appears in the visible answer. That is usually a content-structure problem: the answer is on your page but not extractable cleanly. Tightening the on-page answer fixes it. More on this in our guide to citation readiness.
• Invisible. Neither your domain nor your brand appears for the query at all. That is usually an off-site problem: the third-party pages the engine trusts for that query do not include you. You fix it where the citations come from, not on your own page.

A single blended score cannot tell these two apart, so it cannot tell you which lever to pull. Working query by query can.

The bar is lower than original research

The phrase “unique content” scares people into thinking they need a proprietary study before they can get cited. You do not. Most winnable queries are won by someone who simply did the thing and wrote down the specifics at a resolution nobody else bothered with.

You ran a migration and recorded what actually broke. You priced a job and can show the real number, not a range. You tested two approaches on a real client and have the before and after. None of that is original research in the academic sense. All of it answers a specific query better than a generic paragraph can, because it contains real numbers and real constraints that a model cannot invent.

This is the cheapest moat available, and it overlaps directly with what AI search rewards as first-hand experience. A few examples of the shape:

• “How much does X cost” answered with your actual invoice numbers and what changed the price, not “it depends on your needs.”
• “What happens when you do X” answered with the specific outcome you observed, including the part that went wrong.
• “X versus Y for [specific use case]” answered from a real decision you made, with the reason, instead of a balanced feature grid.
• “Why did X stop working” answered with the exact cause you found, not a list of things to check.

You only need to clear one bar: be more specific and more useful than the current best answer to that one query. Do the thing, then write it down precisely. That is the whole play.

How to find winnable queries you can actually own

Winnable means two things at once: real people ask it, and the current best answer is still generic. Two free sources surface exactly those, and you already have access to both.

1. Google Search Console queries with impressions but low click-through. Open the Performance report, sort by impressions, and look for queries that show often but get almost no clicks. That pattern is real demand the page is failing to satisfy, and the phrasing tends to be conversational, the way people type into an assistant. These are not invented prompts from a tool. They are the queries your audience is already asking, in their own words. Pull the low-click long-tail queries and treat them as your test set.

Two cautions on the GSC pass. First, AI Overview impressions are blended into your Search data and cannot be cleanly separated, so read the trend, not a precise AIO number. Second, if your site is new, GSC will be close to empty and this play has nothing to act on yet. In that case, pivot to entity completeness and category placement until impressions accumulate.

2. Your sales and support logs. These hold queries only your customers ask, phrased the way only your customers phrase them. Read sales call notes, support tickets, and the “quick query before I buy” emails. The queries that come up repeatedly, that you answer the same way every time, are queries with proven demand and almost no good public answer, because the answer currently lives in your team’s heads. Writing one of those down well is often an instant win. Half the strongest pages I have shipped started as a reply I had already typed three times in support.

Combine the two sources and you get a shortlist with both halves of “winnable”: GSC proves the demand is broad, your logs prove the query is specific enough that only you answer it well today. Then sanity-check the current best answer the obvious way: ask the query in an assistant and read what comes back. If the answer is vague, hedged, or wrong, the query is open.

Why real queries beat auto-generated ones

A lot of tooling will happily generate a list of queries for you. The problem is the same one SEO learned the hard way: garbage in, garbage out. Auto-generated queries are guesses at what people might ask. GSC queries and support logs are records of what people actually asked. The phrasing is different, the intent is different, and the phrasing is what the engine matches against.

This is the principle behind how BlueJar builds its prompt set. Instead of inventing prompts, the methodology is built to test the real buyer queries a brand should be visible for, across ChatGPT, Perplexity, Gemini, and Copilot, then shows which exact queries name you, name a competitor, or get answered with no citation at all. Same idea as the manual GSC pass, run at the scale of a full query matrix and across every engine at once. You do the free version first. The tool is worth paying for when you want the coverage and the per-engine breakdown without doing it by hand.

The honest sizing matters here. AI referral traffic is still small. Semrush, looking at billions of visits across 50,000-plus sites, found AI traffic grew 66% in 2025 but still accounts for just 0.14% of total traffic (April 2026). So why bother now? Because the traffic that does arrive converts. Visibility Labs found ChatGPT referrals converted at 1.81% versus 1.39% for non-branded organic across 94 ecommerce brands, about 31% higher (February 2026). Small, high-intent, and compounding. Owning a buyer query now is a cheap option on a channel that is growing fast.

Want to see which buyer queries name you versus a competitor across ChatGPT, Perplexity, Gemini, and Copilot? Run your free AI visibility audit at bluejar.ai. The first analysis is free, and it shows you the exact queries where you are cited, mentioned, or invisible, so you know which ones to write for first.

A query-first content workflow

Here is the reframe as a repeatable loop, the way I actually run it:

• Mine the demand. Pull GSC queries with high impressions and low clicks, plus the recurring queries from sales and support. That is your raw query list.
• Test the current best answer. Ask each query in an assistant. Keep the ones where the answer is generic, hedged, or wrong. Drop the ones already answered well, since those are uphill.
• Answer one query better. Write a focused page or section that answers that single query with real numbers and real constraints, and put the answer high and clean so it is easy to extract.
• Check per engine. Because engines cite different sources, verify across ChatGPT, Perplexity, Gemini, and Copilot, not just one. A win on one is not a win on all.
• Fix the right layer. Cited-but-not-mentioned is an on-page structure fix. Invisible is an off-site placement fix. Match the fix to the failure.

The reframe is small but it changes the whole queue. Stop auditing pages for uniqueness. Start auditing queries for whether the best public answer is still generic. The first test tells you what is polished. The second tells you what you can win. For the bigger picture on how this fits the shift from search engines to answer engines, see our primer on generative engine optimization and the zero-click reality driving it.