Open your keyword tool and look at what you’re targeting: “buy solar leads”, “lead generation agency”, “dead quote follow up”. Two to four words each. Now look at what your buyers actually type into ChatGPT or Google’s AI Mode at 9pm: a full paragraph describing their business, their backlog, their budget and the thing they’ve already tried that didn’t work.
Those two things are not the same query, and a page built for the first one routinely loses the second. This post is about how to optimise for conversational AI search prompts: what long buyer prompts actually contain, how to structure pages that answer them, and how “prompt research” replaces classic keyword research when the query is a paragraph.
The short answer: To optimise for conversational AI search prompts, stop building pages around 2–4 word keywords and start building them around full buyer situations. Long prompts bundle context, constraints, budget and location into one question — so the winning page answers a specific scenario, not a generic topic.
- Google’s own data (May 2026): “the average AI Mode search is triple the length of a traditional Search query”.
- Semrush’s clickstream study (1B+ lines of US data, Oct 2024–Feb 2026): non-search ChatGPT prompts averaged 24.9 words in early 2025; 65–85% of ChatGPT prompts matched no keyword in Semrush’s 27-billion-keyword database.
- For contrast, Backlinko’s analysis of 306M keywords found “The average keyword is 1.9 total words.”
- Structure that wins: an answer capsule up top, scenario headings phrased as real questions, an FAQ that matches long-tail phrasing, and honest constraint-based recommendations.
- Research shifts from “what keywords have volume” to “what would this buyer actually type into ChatGPT”.
The data: your buyers stopped typing keywords
Traditional search is short. Backlinko’s study of 306 million US keywords found the average keyword is just 1.9 words, and while “91.8% of all search queries are long tail keywords”, those long tails “are responsible for a relatively small percentage of total search volume (3.3%)”. Fifteen years of SEO tooling and habit were built on that shape: short head terms carry the volume, so pages target short head terms.
AI search broke the shape. Semrush analysed more than 1 billion lines of US clickstream data covering ChatGPT usage from October 2024 to February 2026. Two findings matter here:
- “The average prompt length for non-search queries nearly halved: from 24.9 words to 13.5 words” — meaning even after settling down, a typical conversational prompt is still roughly 7× the length of Backlinko’s average keyword.
- “The average prompt length for search-enabled queries nearly doubled: from 4.7 words to 8.7 words” — the prompts that trigger live web retrieval (the ones your page can actually get cited in) are getting longer, not shorter.
Maybe the most under-discussed number in that study: for most of the period, between 65% and 85% of ChatGPT prompts couldn’t be matched to any keyword in Semrush’s 27-billion-keyword database. The majority of AI-search demand is invisible to keyword tools. You cannot find these queries in Ahrefs or Semrush’s keyword magic tool, because they’ve never been typed into Google.
Google says the same thing about its own product. In its May 2026 report on AI Mode usage in the US, Google states “the average AI Mode search is triple the length of a traditional Search query”, that planning-related queries “have grown faster than AI Mode queries overall by 80% in the past 6 months”, and that brainstorming queries have grown 30% faster than queries overall since launch. Planning and brainstorming queries are exactly the shape of a buyer describing a situation and asking what to do.
Honesty box — about the famous “23 words” stat. You’ll see AEO roundups claim the average ChatGPT prompt is 23 words versus ~3.4 words for a Google search, with a maximum observed prompt of 2,717 words. Those figures trace back to the early-2025 release of the same Semrush clickstream study. We fetched the live study while writing this post: it has since been updated (April 2026) and no longer carries the 23-word or 2,717-word figures — the current version reports the 24.9 → 13.5 word trend quoted above. The direction of the claim holds (prompts are several times longer than search queries, and the long tail is extreme), but we can’t verify those two exact numbers at the primary source today, so we’re not building on them — and we’d suggest you don’t either.
What’s actually inside a paragraph-length buyer prompt
Long prompts aren’t padded versions of keywords. They’re briefs. Here’s the kind of thing a services buyer types — this is an illustrative prompt we wrote as an example, not user data or a quote:
“We’re a Sydney builder sitting on about 80 unclosed quotes from the last 18 months. We don’t want to hire another salesperson and our admin doesn’t have time to chase people. What’s the best way to reopen those conversations without annoying past prospects, and roughly what should it cost?”
Count the moving parts. That one prompt contains:
- Identity and location — a Sydney builder, so Australian compliance, Australian pricing norms, local examples.
- An asset — 80 unclosed quotes, which makes this a database-reactivation problem, not a new-lead problem.
- Constraints — no new hire, no admin time. Any answer that says “just get your team calling” fails the brief.
- A relationship condition — “without annoying past prospects” rules out spray-and-pray tactics.
- A budget question — the page needs to talk honestly about cost structures, not hide them.
A generic “lead generation services” page answers none of that. The AI engine assembling a response needs passages that address the specific scenario — and it will pull them from whichever site wrote them down. When we ask why a competitor got cited for a buyer prompt and we didn’t, the answer is almost always the same: they had a passage that matched the constraint and we had a page that matched the keyword.
Why this kills keyword-page thinking
Keyword-page thinking says: find the head term, build one comprehensive page, rank it, and let it catch the long tail. That worked because ten blue links let the searcher do the last mile themselves — they’d click your generic page and mentally map it onto their situation.
AI engines do the last mile for the buyer. They retrieve passages, score them against the full prompt — constraints included — and synthesise one answer. There is no click where your generic page gets a chance to be interpreted generously. Either you wrote something that speaks to “80 unclosed quotes, no salesperson, Sydney”, or the engine quotes someone who did.
We covered the mechanical half of this in our post on why your answer needs to sit in the first 30% of the page: retrieval systems chunk pages and lift short spans, so position and self-containment decide whether a passage survives extraction. That post is about where the answer sits. This post is the other half: what questions your pages need to answer in the first place. Get the question wrong and perfect capsule placement just positions the wrong answer beautifully.
How to structure pages that answer paragraph-length prompts
1. Open with an answer capsule for the core scenario
Directly answer the page’s target scenario in 2–4 sentences plus a handful of sourced facts, inside the first screen. The capsule formula (direct, self-contained, evidenced) is in the answer-capsule guide — the job of this page’s method is to make sure the capsule answers a question buyers actually ask.
2. Write scenario headings phrased as real questions
Instead of “Our Database Reactivation Process”, write H2s and H3s the way the prompt is phrased: “What if you have hundreds of old quotes but nobody to chase them?” or “Is it worth following up quotes that are over a year old?”. Each heading plus the paragraphs under it becomes a retrievable unit that matches one slice of the conversational prompt — the constraint, the budget worry, the “for my situation” detail.
3. Build FAQ blocks that match real long-tail phrasing
Not invented FAQs — questions lifted from sales calls, quiz answers, support emails and quote objections, kept in the buyer’s own words (“do I have to lock into a retainer”, not “what are the engagement terms”). Mark them up with FAQPage schema so the phrasing is machine-legible.
4. Make honest, constraint-based recommendations
Long prompts contain disqualifying conditions, and engines increasingly relay them. Pages that say “if you have fewer than a few hundred contacts, database reactivation probably isn’t worth outsourcing — do it manually” earn citations for constrained prompts precisely because they don’t pretend to fit everyone. Honesty is a retrieval strategy, not just an ethical one.
| Keyword-page thinking | Conversational-prompt thinking | |
|---|---|---|
| Unit of research | Keyword + monthly volume | Buyer prompt: situation + constraints + ask |
| Where demand shows up | Keyword tools (Ahrefs, Semrush) | Mostly invisible — 65–85% of ChatGPT prompts match no keyword in Semrush’s database |
| Page target | One head term, broad coverage | One scenario family, specific coverage |
| Headings | Topic labels (“Our Process”) | Questions in buyer phrasing (“What if I can’t hire a salesperson?”) |
| Content shape | Comprehensive, reader does the mapping | Self-contained passages that survive extraction |
| Recommendations | Always “yes, buy this” | Constraint-based, including who it’s wrong for |
| Success metric | Rank position, organic clicks | Share of answer: how often engines cite you for buyer prompts |
Prompt research: the new keyword research
You can’t pull conversational prompts from a keyword tool, so you build the list yourself:
- Mine your own funnel. Sales-call recordings, qualification-quiz answers, email threads and quote objections are transcripts of the exact language buyers will reuse in ChatGPT. The question they asked your closer is the prompt they’ll type when they research your competitor.
- Write the 9pm prompt. For each service and vertical, draft what a real buyer would type into ChatGPT: first person, with their industry, city, backlog, budget worry and the thing they already tried. Aim for 15–30 prompts per service line.
- Poll the engines on those prompts. Run your prompt set through ChatGPT (with search), Gemini and Perplexity on a schedule and record who gets named and cited. This share-of-answer tracking — the methodology we walk through in our guide to getting cited by ChatGPT — is your new rank tracker. We poll our own buyer-prompt set across engines every week; it’s how we know which scenario pages are working within days, not quarters.
- Build one page per prompt family. Cluster prompts that share a scenario (“old quotes, no time to chase” is one family across builders, brokers and gyms) and give each family a dedicated page with the structure above.
Worked examples: buyer prompts and the pages that win them
All prompts below are illustrative examples we wrote to show the pattern — they are not user data, client quotes or statistics.
“Sydney builder, 80 unclosed quotes, no salesperson”
The winning page is a dead-quote follow-up page, not a lead-generation page. It opens with a capsule answering whether old quotes are worth reactivating, has an H3 for “What if nobody on the team has time to make calls?” (the constraint), covers Australian SMS and calling compliance (the location), and talks straight about cost models. It can also carry real reactivation numbers: in our Colliers-era database reactivation campaigns we averaged a 4.4% appointment conversion rate on dormant databases, peaking at 8.9% — the kind of specific, on-topic figure an engine can lift into an answer about whether old quotes are worth chasing.
“US personal-injury firm, intake team misses after-hours calls, no retainer appetite”
The winning page addresses after-hours lead response for law firms specifically, with a section answering “Can we pay per signed case instead of a retainer?” honestly — including when pay-per-result pricing doesn’t make sense. The constraint (“no retainer appetite”) is the heading, not a footnote.
“Gym owner, 300 ex-members, tried a win-back email that flopped”
The buyer has already failed once, so the winning page leads with why single-channel win-backs underperform and what a multi-touch reactivation sequence looks like — answering the prompt’s embedded history (“we already tried X”) rather than pitching the thing they just watched fail.
Where our numbers come from
We’re not writing about conversational prompts from the outside. LeadsNow is a pay-per-result AI lead generation and appointment-setting agency: since 2017 our AI systems have booked 50,769+ sales appointments and generated over 1 million leads, and the buyer conversations behind those numbers are where our prompt research starts. There are 25 filmed client case studies on the site — unscripted clients on camera — and we hold a 4.6-star rating across 43 Google reviews. The reason we care this much about AI-search buyers specifically: AI search leads convert higher than organic — someone who typed an 80-word brief into ChatGPT and clicked through anyway is about as qualified as inbound gets.
Want the buyer-prompt research done for you — with appointments on the end of it? Book a call and we’ll walk through what your buyers are asking AI engines right now, and whether a pay-per-result model fits.
FAQ: optimising for conversational AI search prompts
What is a conversational AI search prompt?
It’s the paragraph-length question a buyer types into ChatGPT, Gemini or Google’s AI Mode instead of a short keyword — typically bundling their situation, location, constraints and budget into a single ask. Optimising for these prompts means building pages that answer specific buyer scenarios rather than generic head terms.
How much longer are AI prompts than normal search queries?
Google’s May 2026 AI Mode report states that “the average AI Mode search is triple the length of a traditional Search query”. Semrush’s clickstream study of 1B+ lines of US data found non-search ChatGPT prompts averaged 24.9 words in early 2025 (settling to 13.5 by early 2026), while prompts that trigger web search nearly doubled from 4.7 to 8.7 words — against an average keyword of just 1.9 words in Backlinko’s 306M-keyword study.
Is keyword research dead, then?
No — it’s demoted. Keywords still show you the language buyers use and the intent behind it, but 65–85% of ChatGPT prompts match no keyword in Semrush’s 27-billion-keyword database, so keyword tools can’t see most AI-search demand. Use keywords as a language input, then rewrite them into full buyer prompts and build pages against those.
How many scenario pages do I need?
One page per prompt family — a cluster of prompts sharing the same core scenario, like “old quotes, nobody to chase them” or “missed after-hours calls”. Most service businesses find 10–20 families cover the bulk of their real buyer prompts. Depth on one scenario beats thin coverage of thirty.
Where do I find real buyer prompts without search-volume data?
From your own funnel: sales-call recordings, quiz and form answers, email threads, and quote objections. Draft first-person prompts in that language, then verify them by polling AI engines on your prompt set and tracking who gets cited — share of answer replaces search volume as the demand signal.
Does this actually produce better leads?
In our experience, yes — a buyer who described their whole situation to an AI engine and still clicked through to you arrives pre-qualified. We’ve broken down the mechanics in why AI search leads convert higher than organic.
See if we’re a fit
Three quick questions. If it’s a fit, our live calendar loads on the next screen. If it isn’t, we’ll point you to free resources instead — you won’t have to sit through a sales call to find out.
Check If You Qualify 👇
We get paid a performance fee equivalent to 10–20% of the sales we help you generate.
Are you OK with that?
If you’re not willing to pay 10–20% as a performance fee, are you happy to pay a $4,000+ per month retainer?
How many leads per month do you currently get?
What’s your current advertising spend or marketing budget (Meta, Google, SEO, etc.)?
What’s the average sale worth to you over that customer’s lifetime?
Given your business currently gets less than 10 leads per month, we’d need to do much more groundwork to set up end-to-end sales systems. Are you OK with a $2,000/mo retainer to do so? (no lock-in)
What’s your work email?
Hey! We might be able to add $100k+ / mo... Enter your email to choose a time!
We’re probably not the right fit — yet
Our model is pay-on-performance — we only win when you’re making sales, and it works best alongside an active marketing engine with advertising budget to get seen. Booking a call now would waste your time, and we’d rather be straight with you.
Grab the free stuff instead — it’s the same playbook we use:
Read the growth blog · Lead-gen FAQ
When the timing’s right, come back — the calendar will be waiting.