Most content teams have already rewritten how they format articles for AI systems — clear headings, direct answers, tidy structure. Far fewer have rewritten how they decide which articles to write in the first place. That second step is still handled the old way: pull a keyword tool, sort by search volume, build the calendar off the top rows.

That habit is now quietly working against you. It filters out exactly the questions your best buyers are asking, for a simple reason — those questions are rarely typed into a search box anymore. They’re spoken or typed to an AI assistant instead, in full sentences, and a volume column has no way to register them.

Why one search became many

The mechanism is worth understanding before the fix makes sense. When someone asks an AI assistant a question, that question isn’t run as a single lookup. Google’s own documentation on AI Overviews and AI Mode describes a “query fan-out” process: the system issues several related searches across subtopics and sources, then assembles a response from what comes back. Google has also said its models keep identifying supporting pages while the answer is still generating, part of why AI-driven results draw from a wider spread of sources than a standard results page. ChatGPT works comparably — OpenAI’s documentation describes rewriting a user’s prompt into retrieval queries before it answers.

Different vendors, same shift: the searching hasn’t disappeared, it’s moved. The model does the query-splitting on the user’s behalf, and the visible prompt is really a bundle of smaller searches stitched into one answer.

That shift is what breaks volume as a way to prioritize content. You’re no longer competing for one keyword’s results page — you’re competing to be one of several sources pulled into a synthesized answer. A page can get cited for a sub-question it never explicitly targeted, or get skipped entirely on a topic it “owns,” because it only ever covered the headline term and never the supporting questions underneath it. This mirrors a pattern we’ve covered before: Google’s own data shows searchers have moved past keywords faster than most content strategies have.

Why keyword tools can’t see the real demand

Keyword tools count strings people type. Prompts don’t resemble those strings. They’re longer and conversational, describing a situation rather than stating a query — “best CRM for small business” versus something closer to “I run a 12-person agency, I’m on spreadsheets now, I need something my team will actually use and I can’t afford a six-month rollout.”

The second version carries real constraints and a decision the person is trying to make. It’s worth considerably more to the business too, because whoever is asking it is closer to acting — and it will show essentially zero volume in any keyword tool, because almost nobody types that exact sentence.

Sort a content calendar by volume and the traditional phrasing wins every time, while the higher-intent version never makes the list. That means prioritizing what’s easiest to count rather than what’s most valuable to answer — a pattern that shows up in why publishing more content is quietly making SEO performance worse when volume rather than value drives the calendar.

Four things to prioritize instead of the volume column

1. Sub-questions over the head term

If a prompt fans out into sub-queries, those sub-queries are the actual targets. Take the head term and list the eight to ten things a buyer needs answered before they can act on it. For “best CRM for small business,” that’s pricing tiers, migration effort, integrations, contract length, and what happens to data on the way out. Then check honestly whether the existing page answers those, or just gestures toward them — most pages do the latter.

2. Entities and concepts over exact strings

Synthesis matches on meaning, not exact phrasing, so repeating a keyword variant buys much less than it used to. What earns a place in an AI-generated answer is covering the subject properly: naming the relevant products, standards, methods and alternatives, and explaining how they relate. That argues for fewer pages built around phrase variants and more pages that cover one subject thoroughly.

3. The decision over the definition

People rarely ask assistants for plain definitions anymore — they get those instantly and move on. What they use assistants for is deciding. Comparisons, selection criteria, trade-offs and objections need to be primary content, not a section bolted onto the end of a guide. A page that can’t support a comparison won’t be much use when a model is answering a comparison prompt.

4. Keep volume where it still governs the decision

None of this retires search volume outright. Plenty of queries remain short, transactional and settled by a normal results page — brand terms, product terms, “near me” searches. Volume is still the right signal there, and there’s no need to rebuild those pages around conversational phrasing. The job is knowing which of your pages belongs to which category, and treating them differently.

Where to find the raw material for prompt-shaped questions

None of this requires new tooling. The inputs are already available:

• People Also Ask boxes and related searches, for the shape of likely sub-questions.
• Reddit, Quora and industry forums, where real people describe their situations in context.
• Sales calls and support tickets — arguably the most underused source, since customers describe problems to a salesperson in almost the same language they’d use with an assistant.
• The assistants themselves. Ask the head term and note which sub-questions the model chooses to address on its own.

Ranking a list that has no numbers on it

This trips up experienced SEOs specifically because it breaks a habit. If a question shows zero volume, it can’t be ranked against one showing 2,400 searches a month on the same scale — but it can be scored on proxies:

• Business value. Does answering it move someone closer to a purchase? A question forty mid-decision buyers are asking beats one four thousand idly curious people are asking.
• Evidence it’s actually being asked. It doesn’t need volume, it needs proof of life — a People Also Ask entry, a live forum thread, the same scenario coming up in three sales calls in a row.
• Coverage gap. Is it already half-answered somewhere on the site? Strengthening an existing page usually beats starting a new one from zero.
• Answerability. Can it be answered concretely, with specifics? Vague questions produce vague content, and vague content rarely gets cited.

It’s slower than sorting a spreadsheet column. It also surfaces demand a volume-only process never would have found, which is the entire point — related to what we’ve written about content strategies that end up accidentally recommending competitors simply because they never got specific enough to be the cited source themselves.

Measuring whether it actually worked

Reporting in this environment tends to boil down to two things worth tracking. First, whether AI-driven sources are sending real visits, and to which pages — that requires analytics configured correctly, since GA4’s default channel grouping tends to split and undercount AI-assistant traffic rather than isolate it. Second, ordinary organic performance: a page that genuinely covers its sub-questions tends to pick up long-tail rankings regardless of whether any single assistant ever cites it. For a deeper look at why standard dashboards understate this shift, see how to diagnose where AI search is skipping your content.

What’s not yet possible is tying one specific citation back to one specific prioritization decision — assistants don’t expose that data. Treat this as a directional program measured over time, not a campaign with clean attribution.

Keyword research isn’t dead — the volume-only filter is

Working out what an audience wants to know, in what order and with what intent, matters more now than when a results page was the only surface being optimized for. A model assembling an answer from multiple sources is far less forgiving of a page that half-covers its subject than a ranking algorithm ever was — a point echoed in why AI search only feels disruptive to sites whose SEO was shallow to begin with.

A practical starting point

Take the ten strongest commercial pages on the site and write out the full prompt a real buyer would use to reach each one — not the keyword, the whole sentence, constraints included. List the sub-questions that prompt would fan out into, and mark which ones the page actually answers. Fix the gaps on the strongest existing pages first, then work through the rest in order of commercial value.

Search volume was always a proxy for demand, and it worked well while demand mostly arrived as typed strings. Now that a growing share of it arrives as a described situation instead, the proxy has stopped tracking the thing it was standing in for. Worth checking before anything else: does the current query list actually sound like the way customers talk?

Frequently asked questions

Why does search volume miss AI-driven demand?

Keyword tools count strings people type into a search box. AI prompts are longer, conversational sentences that describe a situation, and almost no two people phrase the same underlying need identically — so a volume column shows near-zero even when the demand behind it is real and commercially valuable.

Should content teams stop using keyword research entirely?

No. Short, transactional queries — brand terms, product terms, local intent — are still well served by traditional volume data. The change is limiting volume to that category of query rather than letting it decide the entire content calendar.

How can a team prioritize questions that show no search volume?

Score them on proxies instead: how close the question sits to a buying decision, whether there’s evidence it’s actually being asked (forum threads, support tickets, repeated sales-call scenarios), whether an existing page half-answers it already, and whether it can be answered with real specifics rather than generalities.

What is query fan-out?

It’s the technique Google has described for AI Overviews and AI Mode, where a single user question triggers multiple related searches across subtopics and sources before the system assembles one response. OpenAI’s ChatGPT does something functionally similar by rewriting prompts into retrieval queries.