Discussion starter: an AI search question, posted with an answer from Niraj Raut to open the thread. If you have dealt with this on a site, reply with what you saw, especially where it differs.
- AI Search
Do answer-first paragraphs and summary boxes actually increase AI citations, or have you seen them cut organic CTR by giving the whole answer away?
Disclosure: Niraj Raut, who posted this answer, runs the SEO consultancy linked at the end of it.
1 reply
-
adminAdmin
Short answer: there is reasonable correlational evidence that ChatGPT cites early, clearly stated answers more than material further down a page, no controlled public evidence that adding a summary box raises citations in Google’s AI features, and, outside a narrow set of single-fact pages, little reason to blame your own summary for lost clicks. Where clicks fall, the better-documented driver is the results page answering the query itself, through an AI Overview or a snippet, and that happens whether or not your page leads with the answer. What your formatting influences is whether your page is the source that answer is built from.
The distinction that matters is between queries that can be fully answered in a sentence and queries where the honest answer is “it turns on three things”. On the first kind, an answer-first paragraph costs little that the results page has not already taken, and it improves your chance of being cited. On the second, a short answer that names the variables can earn the click, because the searcher can see the page holds the part that applies to them. So the call should follow query type and business model, not a sitewide template.
What the citation evidence shows, and what it does not
Google’s position is that its AI features need no special formatting. Its guide to optimising for generative AI features in Search (May 2026) says foundational SEO still applies and that artificial “chunking” and AI-specific rewording are not needed. That is not a claim that structure is irrelevant, but it does mean a summary box is not a documented signal for AI Overviews or AI Mode.
On the ChatGPT side, a large public dataset is Kevin Indig’s analysis of 1.2 million AI answers and 18,012 verified citations, published in Search Engine Land in February 2026. It matched ChatGPT’s responses to source sentences using embeddings and found that 44.2% of citations came from the first 30% of the content, 31.1% from the middle (30 to 70%) and 24.7% from the final third. Cited content was also twice as likely to include a question mark, with most question-linked citations coming from headings, and it leaned on definitional language (“X is”, “X refers to”) and a high density of named entities.
Relative to the share of text in each band, the opening 30% is over-represented and the middle 40% is under-represented, so the front-loading pattern is real in that data. Three limits stop it being proof that answer-first formatting causes citations:
- It covers ChatGPT only. It says nothing directly about AI Overviews, AI Mode, Copilot or Perplexity.
- It is observational. The opening of a page is where titles, headings and topic sentences cluster, and pages that open with a crisp definition are plausibly written clearly throughout. Position and “the sentence that defines the thing” are hard to separate.
- It describes pages that were cited. It does not show that adding a summary to an uncited page makes it cited.
A plausible mechanism, reasoned from how retrieval systems generally work rather than documented by any platform: answer engines split pages into passages, score them against the query (and, in AI Mode, against related sub-queries, a technique described as query fan-out), then ground each claim in a passage they can attribute. A sentence that names its subject and states the answer without leaning on the previous paragraph is easier to match and safer to quote. That favours self-contained answers wherever they sit, which matters more than whether they sit in a coloured box.
Useful analogy:
An answer engine behaves a little like a journalist on deadline with a stack of printouts. They quote the sentence that states the fact cleanly and stands on its own, not the one that only makes sense after three paragraphs of context. There is no journalist, only passage retrieval and attribution, but the practical effect is similar: extractable sentences get used, and a summary box is only one place to put them.
Where the lost clicks come from
Independent behavioural data on this is scarce, which is why Pew Research Center’s browsing study is still useful. Across 900 US adults in March 2025, users clicked a traditional result in 8% of visits when an AI summary appeared, against 15% when none did, and clicked a link inside the summary in 1% of visits. Users ended their browsing session 26% of the time on a results page with a summary, against 16% without. Question-format searches produced a summary 60% of the time, and searches of ten words or more 53% of the time. It is US data from March 2025, so treat the size of the gap as indicative for other markets and for today’s results pages.
Two things follow. First, the click loss is attached to the results page answering the query, and that answer gets written with or without your page. If you withhold the answer, the summary or snippet is more likely to be built from a competitor’s page, and the loss is likely to remain. Second, a citation is mostly exposure rather than traffic: at a 1% click rate on links inside summaries, AI citations should be valued and measured for visibility and assisted demand, not expected to replace lost clicks.
Answer-first writing can genuinely cost clicks, but in narrower situations than the worry suggests:
- Short-fact queries with no AI feature on the results page. Snippets are drawn mainly from page content, so a crisp opening sentence can become your ordinary snippet and satisfy the searcher in the results. This is where the “giving it away” concern is real, and it bites hardest on pages whose value is a single fact.
- Summaries that make the page redundant. If a 60-word box holds everything a reader needs, the page has an information problem that formatting will not fix.
It can also help clicks. A conditional answer (“most jobs land between A and B; where yours lands turns on C, D and E”) shows competence in the snippet and tells the searcher the page resolves their version of the question.
How query type changes the call
Query type Illustrative query What the results page may do Effect of answer-first What I would do Single fact or definition “what does APR mean” Answer it in full in an AI Overview or snippet Improves the odds of being the cited source; clicks were low either way Answer in the first sentence; judge the page on citations and branded demand, not CTR Conditional answer “kitchen renovation cost” Give ranges; searchers click for their own case Naming the variables up front can support clicks Lead with the range and the factors, with the detail (table, calculator) directly below Procedure “reset a router” Summarise the steps Simple cases resolve on the results page Summarise the steps, then cover failure cases and edge cases in depth Evaluation or comparison “best CRM for a small agency” List options with short reasons A clear verdict is citable; clicks come for the reasoning State the verdict and the criteria early, then show the evidence Task or tool “mortgage repayment calculator” Send people to tools Little effect; the answer is the tool Skip the summary box How to test it on your own site
This is one of the more testable content changes, because opening paragraphs can be added to a template in bulk and AI visibility now has first-party reporting. The design I would use:
- Hypothesis. Adding a 40 to 80 word answer-first paragraph under the H1 (not a separate box) increases AI feature appearances for those pages without reducing clicks relative to control pages.
- Groups. Take one template with many comparable pages, for example 200 how-to or glossary pages. Split them randomly into test and control, stratified by traffic, and change only the test group.
- Google AI visibility. The “Search generative AI performance” report shows impressions for AI Overviews, AI Mode and generative AI features in Discover, filterable by page, country, device and date. It reports impressions only, with no clicks, CTR or queries, and it reached all eligible sites worldwide on 31 August 2026.
- Clicks and CTR. Use the standard Performance report by page, remembering that AI Mode is counted inside the Web totals and cannot be split out as clicks.
- Other engines. Bing Webmaster Tools’ AI Performance report, launched in public preview in February 2026, shows citations by page across Copilot, Bing’s AI summaries and some partners, again without clicks. For ChatGPT and Perplexity, run a fixed prompt set several times per period and track the share of runs that cite the page, because answers can vary between runs.
- Business outcome. Conversions or leads per page. GA4’s “AI Assistant” default channel group (added May 2026) helps, but traffic without a referrer still lands in Direct.
Size the test from the pre-period rather than picking a standard window. Compare the two groups’ AI impressions and clicks for several weeks before the change: if they drift apart by, say, 10% from week to week with no change at all, a 5% effect is unlikely to show in a fortnight. Keep the control group running throughout, because AI feature behaviour shifts with Google’s own model and layout changes, and a simple before-and-after comparison would absorb those. If a core update lands mid-test, extend the test rather than reading the result across it.
What would change the recommendation: if AI impressions rise and the test group’s CTR holds against control, roll the change out to that template. If AI impressions rise but clicks fall faster than in the control group, the decision becomes commercial: acceptable on awareness content, not on pages that pay their way through visits. If nothing moves, the summary is cosmetic, and the effort is better spent on the substance of the pages.
Writing answer-first without giving away the page
- Keep the claim and its qualifier in one sentence. AI systems can lift a single sentence. If the caveat sits in the next paragraph, you risk being quoted without it, which is an accuracy and brand problem as much as an SEO one.
- Name the subject in the answer sentence. “It takes about two weeks” cannot stand alone; “Approval for a [project type] in [area] takes about [time]” can.
- Answer, then show what the answer turns on. The first sentence gives the answer and the second names the variables the page resolves. That is citable and still gives the searcher a reason to click.
- Prefer an opening paragraph to a boxed summary. A box that repeats the introduction adds a duplicate passage. If you use a box, make it carry something the introduction does not, such as a decision table or the figures.
- Know the withholding levers and their cost. The standard preview controls (
nosnippet,data-nosnippetandmax-snippet) also limit what Google’s AI features can show from a page. Search Console’s site-level “Search generative AI control”, rolled out to all websites on 31 August 2026, removes a site’s impressions and traffic from AI Overviews and AI Mode. Google says the setting will not be used as a ranking signal for results outside those AI features. Both cut exposure along with the “giveaway”, so they suit content whose value is destroyed by being summarised, not content that merely feels exposed.
A practical position
If this were my site, I would write answer-first opening paragraphs on pages where the honest answer is conditional, because that framing can serve the citation and the click at the same time. On definitional and single-fact pages I would still answer first, accept that CTR is set mostly by the results page, and judge those pages on citations, branded search and assisted conversions. I would not roll out boxed summaries across every template on the strength of one correlation study. I would test on one template with a control group and the AI performance reports, and let the gap between test and control decide whether it goes further.
Need help with this on your own site?
Niraj Raut works with businesses in Nepal, Australia, the UK, Europe and the US on technical SEO, ecommerce SEO, local SEO and AI search.
Add a reply Cancel reply
Replies are open to approved members. Log in or apply to join.