The follow-up-question test: writing content AI assistants keep coming back to

A page passes the follow-up-question test when it answers the two or three questions a reader would ask right after the one it is titled for — on the same page, in passages that each make sense alone. The reason to care is a number most content advice hasn't caught up with: in Profound's analysis of roughly 730,000 cited ChatGPT conversations from US English users between October and December 2025, 12.6% of opening turns carried a citation, against 3.0% by turn 20. The conversation goes deepest at exactly the point the web gets consulted least.

Reading that chart is what changed how we think about depth while building Contexta. The common advice — answer the follow-up so you get cited a second time — assumes the assistant goes back out to the web for question two. Mostly it doesn't. You are not being re-retrieved on the follow-up; you are being re-read, from whatever it already pulled in. That is a different job for your page, and it changes what "depth" has to mean.

What is the follow-up-question test?

The test is one question you ask of a finished draft: if a reader accepted this answer, what would they ask next — and is that answer on this page? Write down the next two or three questions honestly, without steering them toward what you already wrote, then check whether each has a passage that answers it completely in its first sentence or two. Anything you have to say "well, that's a different article" about is a gap, not a link opportunity.

The discipline is in the honesty of the list. The failure mode we hit constantly on our own drafts is inventing follow-ups that flatter the piece — the questions whose answers we'd already written — instead of the awkward ones a real buyer asks, which tend to be about cost, failure, and what happens when the thing doesn't work.

Each child on that tree is a passage, not a paragraph of hedging. If the answer to "can you wash it" is two sentences under a heading that asks it in the reader's words, an assistant can lift it whole — the mechanic behind question headings and why they win snippets and AI answers.

Why does answering the next question matter more than ranking for it?

Because on the follow-up turn there is often no ranking involved at all — the assistant answers from what it already retrieved rather than searching again. Profound's turn-by-turn data shows citation rates falling away steadily as a conversation continues, and their explanation is that opening questions need factual grounding while later turns are clarifications and deeper dives that the model handles without fresh web data.

Be careful about how far you push that. Profound's post reports the pattern and offers an explanation for it; it does not document the retrieval mechanism, and neither OpenAI nor anyone else publishes exactly when ChatGPT decides to search again mid-conversation. What the numbers support is narrower and still decisive: a citation on turn 5 is roughly half as likely as on turn 1, so a page that only answers the opening question has one shot at being useful and then goes quiet. In a world where under a third of searches end in a click, staying inside the answer for the whole conversation is most of the value you can still capture.

How is Google AI Mode different, and why does that change nothing?

Google inverts the timing but lands on the same requirement: AI Mode asks the follow-ups itself, up front, before the user ever types them. Google describes the mechanism plainly — "AI Mode uses our query fan-out technique, breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf" (blog.google, May 20, 2025, checked July 2026). One question becomes a spread of related searches, resolved in parallel, then synthesised.

So the two big surfaces fail your shallow page for opposite reasons. Google fans out to the subtopics immediately and finds someone else's page for each one you didn't cover. ChatGPT never fans out, and simply has nothing of yours to work with when the reader digs in. Either way the page that answers the neighbourhood of questions is the one that survives, which is why the test is worth running even though the two engines behave nothing alike under the hood.

How do you find the follow-up questions without guessing?

Open Google Search Console, filter to the single URL, and read its query list — the follow-ups are already there, as the long-tail queries that page collects impressions for and rarely converts. This is the part people skip. Brainstormed follow-ups reflect what the writer finds interesting; the query list reflects what people actually typed after landing on that topic, including the phrasings you would never have guessed.

Two practical notes from doing this on real stores. Queries with impressions but near-zero clicks on a page you already own are the strongest signal — the page surfaced for that question and did not satisfy it, which is precisely a missing follow-up. And low-volume queries matter more here than in classic keyword work, because a conversational assistant is matching meaning, not volume, so a question asked forty times a month is a real passage worth writing.

Turning that list into passages is where Contexta's AI editor does the work: it pulls a page's own Search Console queries and drafts answer-first sections against them in the site's learned voice, so the follow-ups you cover are the ones the data says people asked rather than the ones that were convenient to write. The judgement about which questions deserve a passage stays yours — the tool removes the excuse that the list was too long to work through.

How many follow-ups should one page answer before you split it?

Cover the follow-ups that share the original question's intent, and split when a follow-up would need a different page to satisfy someone arriving cold. "How long does the battery last" belongs with "are they waterproof" because the same shopper asks both in one sitting. "How do lithium battery import rules work" does not — a reader searching that has a different job to do and will not want the jacket page.

The trap on the other side is padding, and it is worth naming because the follow-up test is easy to abuse into word count. Adding sections that answer nothing new does not deepen a page; it dilutes the passages that were working, since a retriever scoring your page for one question now has more competing text to wade through. We go through the actual relationship between length and passage quality in content length and how AI search reads passages — the short version is that the number of self-contained answers is what matters, not the number of words wrapped around them.

Is this just "write in-depth content" with a new name?

Largely, yes — and that is the honest answer rather than a reason to skip it. Google's own AI optimization guidance is blunt that generative features run on the same machinery as the rest of Search: "The best practices for SEO continue to be relevant because our generative AI features on Google Search are rooted in our core Search ranking and quality systems" (developers.google.com, checked July 2026). The same guide tells site owners there is no special markup, no chunking, and no rewriting-for-AI step required. Anyone selling you an AI-specific content format is selling you something.

What the follow-up test adds is not a new discipline but a check that catches a specific, common failure: content that is long and still shallow. A 2,500-word page can answer exactly one question and pad the rest, and until you write out the reader's next three questions you generally cannot see that from inside the draft. It is the same instinct behind whether FAQs still matter in 2026 — not the schema, not the accordion, but the question of whether a real question is getting a real answer.

Run it as the last step before publishing, on the finished piece rather than the outline, and expect to fail it more often than you would like. We do.