AI Mode Follow-up Questions Content Strategy: What Smart SEOs Do – AI Mode Follow-Ups Are the New Keyword Research
When Google AI Mode logs a follow-up question as a query, it hands you something keyword research never could: the question a user asks after they’ve seen an answer.
On one page I classified 48 of them. Together they pulled 308 impressions and 3 clicks – worthless as traffic, and the clearest content brief in the whole client account.
Each one is a follow-up in the conversation, and every follow-up is a decision about whether you keep the citation or hand it to a competitor. This is how to read them, and what they mean for AI SEO strategy.
By Konrad Szymaniak
28 Aug 2026 · 12 min read
Keyword research has always had a blind spot, and nobody could do anything about it. It tells you what a person types to start… the cold query, the first move. It has never told you what they ask next. What they say once they’ve read the answer and want more, or want it in a table, or don’t believe it.
That second move… post-answer intent… was invisible. It happened in someone’s head, or on a competitor’s page, and you never saw it.
Google AI Mode just made it visible for me to analyse AI Mode Follow-up Questions and create a smarter Content Strategy
Because AI Mode counts every follow-up as a brand-new query and logs it in the standard performance report…. impressions, position, the lot, the report now contains a readable stream of what users ask after your page has already appeared in an answer. John Mueller confirmed the mechanism in August. The dedicated Generative AI report Google launched on 3 June strips the queries out; the old performance report leaks them. So the follow-ups are sitting in an account you already own, and almost nobody is reading them as what they are.
I took one page…. anonymised here (client data); the niche is UK medical admissions, the kind of page anxious applicants argue with an AI at 4am – and pulled every follow-up turn attributed to it over 28 days. Then I classified them using a regex. Here’s what that set is worth, and why it changes how you brief content.
48 follow-ups, 308 impressions, 3 clicks, 28 days
That’s the headline, and the traffic numbers are the point because they’re so bad. Forty-eight distinct follow-up turns. 308 impressions between them. Three clicks – all three from a single row, “what about cambridge”. Everything else: zero.
If you model these rows statistically the way you model a query report – by clicks, by CTR – you delete the lot. They’re the definition of a low-value long tail. A rank tracker ignores them. A weekly report never surfaces them.
And they are the most valuable rows in the client account, because they aren’t traffic. They’re intelligence. Thirty of the forty-eight are pivots – “what about…”, “how about…” – and those thirty carry 254 of the 308 impressions.
Eighty-two per cent of this signal is the user redirecting the conversation.
Each conversation change is a real customer, mid-thread, telling you exactly where their attention went the moment after they got an answer that involved your page. You cannot buy that data. No tool sells it. It’s specific to the conversations your content is already inside.
Every pivot is a customer follow-up
Here’s the mental model that makes this actionable. A follow-up like “what about oxbridge” or “what about international students” is not a keyword to rank for. It’s a follow-up in the conversation – a point where the thread tries to go somewhere new. And at every follow-up, one of two things happens.
If your page covers where the thread wants to go, Google keeps citing you, and you hold the conversation into the next turn. If it doesn’t, the follow-up routes to whoever does cover it, and you’ve been dropped mid-conversation. Not because you ranked badly for a keyword – because the thread asked a question your page couldn’t hold.
So the follow-up list isn’t a curiosity. It’s a prioritised content roadmap written by your users’ actual AI conversations.
The impressions tell you which follow-ups matter. The position tells you which ones you’re already losing. That is a brief no third-party keyword tool has ever been able to produce, because keyword tools model the first move and this is the fifth.
The map: cluster the follow-ups and the page brief writes itself
Sort those follow-ups into themes and the conversation hands you its own outline. From this one page:
| Follow-up theme | Impr | What the thread is asking for | The build it implies |
|---|---|---|---|
| Comparison (Oxbridge / “both” / “or oxford”) | 27 | Set this option against the obvious rival | A real comparison block, not a mention |
| Reassurance / gut-check (“is it competitive”, “is it prestigious”, “does it actually matter”) | 9 | Tell me honestly how hard this is | An honest “how competitive is this really” section |
| Adjacent scope (“what about gcses”, “english literature”, “do they do medicine”) | 7 | The requirement next to the one I read | Coverage of the neighbouring requirements |
| Format command (“put that in a table”, “give me the requirements”) | 5 | Restructure the answer | Data as clean tables the model can lift |
| Score-test (“how about 2070”, “what about 1800”) | 3 | Judge my exact number against the bar | A score-band table with a verdict |
| Segment (“international students”, “is that international”) | 4 | The version of this for my situation | An international-applicant sub-section |
| College-level (“jesus college”, “other colleges”) | 2 | One level of granularity below the page | Sub-entity detail, or a clear “see college pages” |
The comparison follow-up is the single biggest themed pivot on the page – the thread keeps trying to weigh this against Oxford, and if the page doesn’t handle the comparison, that’s 27 impressions of conversation walking out the door.
The reassurance cluster is a whole emotional register the page probably answers in facts when it needs to answer in judgement.
And the score-test follow-up is the highest-leverage build in the table: users are pasting exact numbers to be graded, and a page with a proper score-band verdict captures a class of turns almost nobody answers.
The one row your instinct should catch: “put that in a table” is a High-confidence prompt – a person telling a machine to reformat. It converts nobody. But it’s a structural instruction…. and your content should match this.
If the model is being asked to table your content, then serving that content already tabled -clean, dated, extractable – is how you make yourself the easiest source to lift. The command is a hint about form, not just topic.
Position tells you which follow-ups you’re already losing
Of the 48 follow-ups, 28 rank in the top 3.
Seven rank beyond position 5 – this is where you’re being dropped.
Read the drop-offs first, because they’re where the conversation is leaving you at high demand.
The biggest leak on this page: “how about medicine” – 13 impressions, position 5.2. A high-demand pivot where the page is slipping out of the answer. Then the comparison forks: “what about oxbridge?” at 5.7, “what about oxford and cambridge?” at 7.7 – the thread pivots to comparison and the page can’t hold it, exactly the gap the map predicted. “what about at cambridge” sits at 9.5, “what about in cambridge” at 12.
PSSS: This is where other pages on the client’s account are being cited above the current one.
But, if you dont have these pages then… that’s a to-do list with the priority already scored: fix the coverage the high-impression, high-position follow-ups point to, re-pull in a month, and watch the position on those follow-ups fall toward 1. It’s a conversation follow-up optimisation loop with a feedback signal SEO has never had. Basically, you’re watching whether the conversation stopped leaving.
The thread is crossing two of your pages
One more finding, and it’s the one with the biggest structural implication. The follow-ups don’t all resolve to a single URL. Five of them are attributed to a second page on the same site (as mentioned above), a narrower, graduate-entry page – while the rest sit on the broad requirements page.
On that second page, “how would you get in” ranks position 1 and “is it prestigious?” ranks position 1. Google is routing specific follow-ups of one conversation to specific pages on your domain – the broad page holds the requirements follow-ups, the narrower page holds the process-and-prestige follow-ups.
That is a strong position to be in and a decision you have to make.
Google is willing to keep a conversation on your site across multiple follow-ups… but only if the pages are clearly differentiated and interlinked, so the model can move between them instead of leaving. If your two pages overlap and compete, you’re splitting your own citation authority and inviting Google to cite a cleaner competitor. So the strategic call is hub-and-spoke: decide which page is the hub for the conversation, make the pages genuinely distinct, and link them so the model can cite the conversation across your domain. The follow-up data is telling you the site architecture the conversation wants.
The strategic implications
Pulling it together, this is what changes once you start reading follow-ups as a feed rather than a curiosity.
1. Content strategy gains a second intent layer. Keyword research models pre-answer intent — the cold first move. Follow-ups model post-answer intent — the second, third, fourth move. You now brief pages against both: the query that earns the first citation, and the follow-ups that decide whether you keep it. Most teams optimise only the first and wonder why they get surfaced once and dropped.
2. The unit of optimisation is the conversation, not the keyword. You’re no longer trying to rank a page for a keyword. You’re trying to build a page that holds a conversation as it moves. That means coverage breadth (answer the likely follow-ups on-page) and structural clarity (tables, dated facts, shown working) so the model keeps choosing you turn after turn.
3. Anticipating the follow-up is the whole job. Every “what about X” is a branch. Cover X and the conversation stays; don’t and it walks away. Map the follow-ups, weight them by impressions, fill the gaps in order. Your users have written your content roadmap; you just have to read it.
4. Measurement inverts — stop grading follow-ups on CTR. 308 impressions, 3 clicks. As traffic rows, a rounding error. As intelligence, the best input in the account. The metrics that matter are breadth – how many distinct follow-up turns you appear for, i.e. how much of the conversation you hold – and depth – your average position on them, i.e. how firmly. Report those two. CTR as a KPI on a follow-up is an error. It just doesn’t make sense.
5. This is proprietary, defensible intelligence. Every competitor in your niche uses the same keyword tools and sees the same volumes. Nobody else can see the follow-ups your pages are in. This feed is unique to the conversations your content already earns, which means acting on it builds you a lead on rivals.
6. Weight the roadmap by confidence, not just volume. These turns aren’t labelled by Google – the classification is from a custom regex. A prompt like “put that in a table” is unmistakable; a bare “what about cambridge” is probably conversational but could be a typed search; “yes both” or “give me the requirements” are lower-confidence guesses. Build against confidence × impressions, and treat the low-confidence, single-impression rows as hints, never mandates.
How to run it
- Open Performance → Connect to API. Isolate the follow-up questions by shape, or with a regex in the query filter.
- Classify by confidence so you know which rows to trust, and set aside the searches from the conversational turns.
- Cluster the follow-ups into themes. The themes are your on-page sections. The impressions are your priority order.
- Read the position column for leaks – every follow-up above position 5 is a follow-up you’re losing at demand. Fix the coverage those follow-ups point to first.
- Check whether the conversation crosses pages, and if it does, make the hub-and-spoke call: differentiate and interlink, don’t let your own pages compete for the same conversation.
- Re-pull monthly and grade on breadth and position, not clicks. The follow-ups you built for should climb; the leaks should close.
A word of caution
Google doesn’t label these follow-ups, so the classification is read from my regex. With some medium-confidence rows. This will be ordinary short searches that happen to look conversational. Single-impression rows are noise on their own; they’re signal only in aggregate, as a cluster. And this is one page, one 28-day pull, in a vertical where the conversational effect runs hot – the volumes will be smaller and calmer on a B2B page. Confirm a follow-up is real, and recurring, before you build a section for it. The pattern is what travels: users’ next moves, logged as queries, clustering into themes, with position marking where you hold the thread and where you drop it.
But the shift underneath is real and it isn’t going back.
For the first time, the question a user asks after they’ve seen your answer is written down, in an account you already own, for free. That’s the input your content strategy has been missing since search began.
Turning a follow-up feed into a content roadmap – clustering the follow-ups, closing the leaks, and rebuilding pages to hold the conversation across more follow-ups – is the work my consultancy, Szymaniak Digital, does.
