Guide · Updated August 2026

AI Overviews Are Eating Your Blog Traffic. Here's What Still Worksand what to stop writing

The click loss is real and it is measured. But the popular fix — “stop writing tutorials, write comparison pages” — is contradicted by the best data available, and the thing that actually changes your outcome is narrower and harder than either.

Updated 19 min readBy the ParseStream team
Start free trialSee multi-platform monitoring
Under a heading reading “ONE INFORMATIONAL QUERY” and a note reading “EACH STEP DIMS”, four cards step down and to the right, each one fainter than the one above it: “the query · ten results, no summary”, “the summary · assembled from several sources”, “the answer is complete · nothing left to look up”, and “no click · the session ends on the search page”. A dashed hairline runs from the right edge of each of the first three cards out to the frame, leaving three descending levels across the empty right-hand margin. A rule closes the staircase off underneath, and below it a raised strip outlined in violet is headed “WHAT THE ANSWER QUOTES INSTEAD” beside a pill reading “cited · not clicked”. Inside the strip, a Reddit icon sits beside the quoted sentence “we ran it for eight months and the export broke twice”, over the line “a thread, not a landing page — the answer lifts the sentence and carries the name with it”.
The short answer

Are AI Overviews really killing blog traffic, and what still works?

Yes, on the queries they appear on. Pew Research Center tracked the browsing of 900 U.S. adults and found people clicked a search result on 8% of visits to a Google page carrying an AI summary, against 15% of visits to pages without one. Ahrefs, comparing 300,000 keywords in Google Search Console, puts the position-one click-through drop at about 58% versus a forecast without AI Overviews.

What still works is not a content format. Seer Interactive's 2026 study found comparison queries trigger an AI Overview 95.4% of the time — 267 of the 280 comparison queries it measured, and more than any other format — so there is no page type to hide behind. What moves the number is being cited inside the answer, worth somewhere between about 29% and 120% more organic clicks per impression than sitting on the same page uncited, depending on how you treat one dominant account in Seer's uncited segment. Either way it is the largest lever on the page. Community threads on Reddit, Hacker News and Quora are among the most-cited sources AI answers pull from, which makes being genuinely present in them the most durable thing left to invest in.

8% vs 15%

share of Google search visits where the user clicked a result link — with an AI summary on the page, and without one.

Source: Pew Research Center
95.4%

of “X vs Y” comparison queries returned an AI Overview — 267 of the 280 comparison queries Seer measured, and the highest rate of any format tested.

Source: Seer Interactive
+29–120%

more organic clicks per impression when your domain is cited inside the AI Overview, versus sitting on the same page uncited. The headline +120% is Seer's raw aggregate; one account is 47% of the uncited segment's impressions at 0.20% CTR, and excluding it the lift is about +29%.

Source: Seer Interactive
01

How much traffic are AI Overviews actually taking?

Enough that the honest answer is “a lot, on the queries they appear on” — and the reason to be careful about the number you quote is that the loudest figures in circulation are the least defensible ones. Three independent measurements agree on the direction and disagree on the size, because they are measuring different things.

The three measurements worth knowing

Pew Research Center is the one with a real-user panel behind it. Researchers tracked the browsing of 900 U.S. adults through March 2025 — 68,879 Google searches, 12,593 of which produced an AI summary — and found that people clicked a traditional search result on 8% of visits to pages with a summary and 15% of visits to pages without. A click on a link inside the summary itself happened on 1% of visits. Sessions ended outright on 26% of pages with a summary against 16% without.

Ahrefs measured the same effect from the publisher side. Comparing 300,000 keywords in aggregated Google Search Console data — 150,000 with an AI Overview, 150,000 informational keywords without — position-one click-through fell from 0.073 (7.3%) in December 2023 to 0.016 (1.6%) in December 2025, against a forecast of 0.037 (3.7%) had AI Overviews never rolled out. That is a roughly 58% reduction, and it is a sharp increase on the same study's April 2025 finding of 34.5%. Ahrefs notes its own caveat: the update spans two years where the original spanned one.

Seer Interactive tracked 53 brands, 5.47 million queries and 2.43 billion organic impressions from January 2025 to February 2026 — and found the picture stopped getting worse. Organic click-through on AI Overview queries bottomed out at 1.3% in December 2025 and recovered to 2.4% by February 2026. Anyone telling you the line only goes down has not read the most recent data.

What Google says, and why both things can be true

Google's public position, from Liz Reid in August 2025, is that “total organic click volume from Google Search to websites has been relatively stable year-over-year” and that “average click quality has increased” — explicitly contrasted with “third-party reports that inaccurately suggest dramatic declines in aggregate traffic.” The post publishes no figures, no sample and no methodology, which is worth stating plainly rather than treating as evidence either way.

The two positions are less contradictory than they look, and the distinction is the most useful thing on this page. Google is talking about aggregate volume across all of Search. The studies are talking about per-query click-through on the subset of queries where an AI Overview appears. Both can hold at once if search volume grows while individual informational queries convert to clicks less often — and Seer's data shows exactly that shape, with click-through on queries that get no AI Overview rising from 2.93% to 3.97% across 2025 even as AI Overview queries collapsed.

The number you should not repeat

A figure in the high eighties — “87% of clicks lost” and variants — circulates widely and traces to no primary source. The nearest arithmetic that produces it divides two Pew numbers that are not a before-and-after pair: the 8% who clicked a result link and the 1% who clicked a link inside the summary, measured simultaneously on the same page views. If you are going to argue about this in public, use the 8%-versus-15% comparison or Ahrefs' 58%, and say which one you mean.

02

Which posts get cannibalised first?

The intuitive answer is right about the mechanism and wrong about the boundary. AI Overviews are overwhelmingly an informational-query phenomenon: Ahrefs found 97.70% of US keywords triggering one were informational in intent, and Seer measured AI Overviews on 36% of informational queries against 8% of commercial and 5% of transactional ones. If your blog is built on definitions, explainers and how-tos, that is where the loss lands first, and Pew's own findings show why — 60% of queries beginning with a question word produced an AI summary, and the rate climbed from 8% on one- and two-word searches to 53% on searches of ten words or more. The longer and more question-shaped the query, the more likely Google answers it itself.

Then the boundary moves, which is the part most versions of this advice get wrong.

AI Overview trigger rate by query format, from Seer Interactive's query-format analysis (February 2026). Each rate is scored on the queries matching that format, not on the 49,353 the format analysis covers in total — the comparison row is 267 of 280. The formats a content team would intuitively treat as safe are at the top.
Query formatShowed an AI OverviewWhat that means for the page you wrote
Comparison — “X vs Y”95.4%Almost always summarised. The page's value is now citation, not the click.
Review queries86.3%Same. First-hand detail is what distinguishes yours from the summary.
Question — what / why / how / is / are85.9%The classic blog post shape, and within ten points of the comparison pages people are told to switch to.
Price / cost / buy83.4%Higher than most people assume. Commercial intent is not a shield.
“Best of” lists81.3%Heavily summarised; the listicle format is not differentiating.
Single-word informational27.3%The lowest rate measured — but still better than one in four.
AI Overview trigger rate by query format, from Seer Interactive's query-format analysis (February 2026). Each rate is scored on the queries matching that format, not on the 49,353 the format analysis covers in total — the comparison row is 267 of 280. The formats a content team would intuitively treat as safe are at the top. Seer Interactive, “AIO Impact on Google CTR: 2026 Update”, 24 April 2026 — 53 brands, 5.47M tracked queries, 2.43B organic impressions. The query-format analysis covers 49,353 distinct queries in total, but each rate above is the share within that one format — the 95.4% is 267 of 280 comparison queries, so it is a confident ordering on a modest base rather than a rate measured across the whole set. Seer's separate intent-based figures (informational 36%, commercial 8%, transactional 5%) classify queries by SERP feature rather than by wording, which is why the two sets of numbers describe the same SERPs differently. Seer states it cannot claim causation. Figures move; re-check before quoting.

“Write comparison pages instead” is not the escape hatch

Seer's 2026 study broke 49,353 queries down by format rather than by intent, and comparison queries — the literal “X vs Y” shape — returned an AI Overview 95.4% of the time. Read the denominator before you quote that: it is 267 of the 280 queries matching the comparison shape, not a share of the 49,353. Each format in the table is scored on its own subset, and the comparison subset is a small one. The ordering still holds — it is the highest rate of any format tested, ahead of review queries at 86.3% and question-format queries at 85.9% — and Seer's own recommendation is blunt: comparison and question-format pages need a citation audit now, not next year.

Semrush found the same drift from a different dataset. Across 600,000 keywords over six months to April 2026, AI Overviews on commercial-intent SERPs grew 71% — with Finance up 231% — while transactional SERPs went the other way, down 5%. Semrush's own example of a commercial-intent query is “Peet's cold brew vs Starbucks cold brew.” The comparison page is not a shelter. It is the front line.

What is genuinely less exposed

Transactional queries, for now. Someone searching to buy, book or find a nearby branch gets a Shopping box or a map, not a synthesis, and both Seer and Semrush have that segment flat or falling. Navigational queries — people typing your brand name — are similarly untouched, which is one reason brand strength has become a search asset rather than a marketing abstraction.

Beyond that, the honest read is that nothing is structurally safe and the exposure is a spectrum rather than a wall. Seer found even single-word informational queries triggering an AI Overview 27.3% of the time across 25 million impressions, and wrote that “the idea that broad, short queries are less AIO-affected is not supported by this study.” Plan for coverage to widen, not for a category to hold.

03

What still works, and why community threads are the odd one out

If no format is safe, the strategy cannot be a format. What the data actually supports is narrower: on a query where an AI Overview appears, the only variable you influence is whether you are inside it.

Seer measured that gap directly, and the honest version of it is a range rather than a headline. The raw aggregate is about 120% more organic clicks per impression when cited than when sitting on the same result page uncited — a full-year informational average of 2.07% click-through against 0.94%. But Seer flags in the same study that the 0.94% is “materially influenced by one account representing 47% of that segment's impressions at 0.20% CTR”, and that excluding that account the uncited aggregate is approximately 1.61%. Against 1.61%, the same 2.07% is a lift of about 29%, not 120%. This page quotes the range for the same reason it refuses the “87% of clicks lost” figure: one dominant account moving an average is exactly the kind of detail the sources that repeat +120% have dropped.

What survives either treatment is the direction and the ordering, which is what a content decision actually turns on. Citation beats non-citation on the same query by a margin nobody's caveat erases. It is still not a return to the old baseline — a cited page underperforms the same query with no AI Overview on it by roughly 38% — so the choice on an AI Overview query was never between citation and the good old days. It is between citation and the bottom of that range.

Why forums are the format AI answers amplify

Pew's analysis of what AI summaries actually link to found Wikipedia, YouTube and Reddit to be the three most commonly linked sources, together accounting for 15% of the sources listed in the summaries examined. Profound's larger sample — more than 4 billion AI citations across 300 million answer-engine responses between August 2024 and October 2025 — puts Reddit first overall at 3.11% of all citations, ahead of YouTube at 2.13% and Wikipedia at 1.35%.

The mechanism is not mysterious. A generated answer needs specific, attributable, first-hand statements, and a marketing page is written to avoid making any. A thread where somebody says what broke, at what scale, and what they switched to is exactly the kind of passage a retrieval system can lift and attribute. That is why community presence behaves differently from every other channel here: AI search does not replace a forum thread, it quotes one — and the quote carries whoever wrote the useful comment inside it.

Free tools, product pages and the shape of AI referral traffic

The other surviving category is the one nobody writes blog posts about. Ahrefs published its own analytics in June 2025 and found more than 80% of its AI-search referral traffic went to three page types — free tools, product pages and its homepage — with free tools alone at 36.4%. That is one company's data about one company, and it should be read as an illustration rather than a benchmark.

It is also worth quoting the part of that post which cuts against the easy conclusion. Ahrefs wrote that “everyone thinks top of funnel content is dead in AI search” but that they were “seeing better performance for all of our informational content types” as a share of AI referrals. The reasonable reading is not that informational content is fine. It is that AI referral volume is small enough that share-of-referrals arguments are fragile in both directions, and that a tool somebody can use is harder to summarise away than a paragraph explaining the same idea.

04

Triaging the blog you already have: update, consolidate, or stop

Most of the advice in this category is about what to publish next, which is the easy half. The harder and more valuable exercise is deciding what to do with the eighty posts already sitting there. Work through it in one pass, and be willing to reach the answer “stop.”

  1. Step 1 — Sort every post by query format, not by traffic

    Pull your top pages from Search Console and label each by the shape of the query it wins: definitional, how-to, question, comparison, best-of, or branded. The format is what predicts exposure, and the table above is your rate card. A post that still gets traffic on a question-format query is not safe, it is early.

  2. Step 2 — Check whether an AI Overview is already there

    Run the actual queries. Note whether an AI Overview appears, and whether your domain is one of the links inside it. This is the single most decision-relevant thing you can gather, because it splits your library into three groups with three different answers: no AI Overview yet, AI Overview with you cited, and AI Overview without you.

    Google Search Console will not do this for you — it does not separate AI Overview impressions and clicks from the rest of your organic data, which is a real reporting gap and the reason this step is manual.

  3. Step 3 — Update the pages where you appear but are not cited

    These are the highest-return edits, because the gap between cited and uncited is the largest measured effect on the page. The changes that help are unglamorous: answer the question outright in the first two sentences, put a specific number or a first-hand detail into the passage a model would lift, and remove the throat-clearing paragraph that currently occupies the position a retrieval system reads first.

  4. Step 4 — Consolidate the thin cluster into one page worth citing

    Most blogs carry six near-identical posts about one topic, written for six keyword variants that no longer behave differently. An AI Overview does not care that you have six; it will summarise the topic once. Merge them into a single deeper page, redirect the rest, and spend the saved effort on making the survivor specific enough to quote.

  5. Step 5 — Stop writing the ones the answer already covers

    There is a genuine category here: short definitional posts, glossary entries and “what is X” explainers targeting queries that now resolve entirely inside the summary. Pew found the median AI summary was 67 words — long enough to finish those questions completely. Writing more of them is not a traffic strategy, and the honest move is to redeploy the budget rather than to optimise a page whose query no longer produces a visit.

    Keep writing the ones where your first-hand experience is the content. That distinction — can a model produce this from what it already has, or does it need you to have done something — is a better filter than any keyword tool.

05

The community-presence playbook, as a channel rather than a tactic

If the surviving asset is a citation, and the most-cited sources are places where people answer each other, then the replacement channel is participation in those places. This is slower than publishing and it does not scale by volume, which is exactly why it holds up.

Being plain about our own position here: this argument is the strategy this site visibly runs. The comparison pages and the guides you are reading are the publishing half of it; the three steps below are the community half, and ParseStream is the product we built because we needed the monitoring layer for them.

Find the conversations before deciding what to say

The unit of work is a thread, not a post, and the threads that matter are the ones where somebody is asking the question your product answers. Track the phrasings people actually use — “any tool for”, “alternative to”, “is X worth it”, “how do you handle” — across Reddit, X, LinkedIn, Quora and Hacker News, which are the five platforms ParseStream monitors. Our guide to choosing social listening keywords covers how to build that list without drowning in noise.

ParseStream's AI Filter applies your own plain-English relevance rule to every incoming mention, Lead Detection judges each one against your brand context, and Intent Detection tags it — buy intent, product question, competitor complaint, pain point, testimonial or promotional. For this channel, product questions and comparison threads are the queue to work, because they are phrased the way people phrase prompts.

Answer as somebody who has done the thing

The comment that gets quoted is the one carrying a number, a constraint or a failure. “It handles high volume” is unquotable. “We pushed about 40,000 events a day through it and hit one rate-limit issue” is quotable, checkable and useful, and it is also the comment humans upvote — which is what puts it near the top of the thread, which is what makes it the passage that gets retrieved.

ParseStream drafts a Suggested Reply for each mention worth answering. You edit it into your own voice and post it yourself; auto-posting is disabled at the server level and always has been. That is not a missing feature on a page like this — a removed comment is cited by nothing, and communities remove obviously automated ones.

Measure it as a channel with a long payback

Nothing reports back from a citation. There is no console, no referral tag, and frequently no click at all, so the honest measurement is a combination of leading indicators: comments that earned real upvotes, the same prompts re-run monthly against ChatGPT, Perplexity and Google AI Mode to see whether you appear, and branded search volume, which tends to rise before anything else does. Treat it the way you would treat a podcast or a conference talk, not the way you would treat a paid campaign.

06

What this argument cannot do for you

Three limits, stated because a page built on other people's studies should be clear about where they stop.

The measurements are correlational. Seer says so explicitly — it cannot claim causation, and notes that higher-authority brands are also more likely to be cited, so some of the citation premium is brand strength rather than a technique anyone can copy. The strongest causal work in this area is a randomised field experiment which we were unable to render-verify from its publisher, so it is not cited here at all.

The figures move fast, in both directions. Seer's own 2026 update forecast continued decline and then recorded a recovery instead; Semrush found AI Overviews expanding into commercial queries while retreating from transactional ones. Any content plan that depends on a specific percentage holding for a year is a plan with a short shelf life. What has been stable across every study since 2024 is the ordering — informational exposure is highest, citation beats non-citation, and community sources are near the top of what gets quoted.

And none of this makes a citation buyable. There is no submission endpoint, no schema that compels a model to quote you, and nobody selling guaranteed AI visibility can deliver it. The only durable position is having said something specific and useful in a place people read, which is the same advice as it was before, now with a worse alternative.

Frequently Asked Questions

How much traffic do AI Overviews actually cost a blog?

Is the “87% of clicks lost” figure real?

Should I switch from writing tutorials to writing comparison pages?

Does being cited in an AI Overview actually recover the traffic?

Why would Reddit and Hacker News threads survive when blog posts don't?

Google says traffic is stable. Who is right?

Related reading

The how-to that follows this argument, and the platforms where the citable conversations happen.

GEO & AEO: how to get cited by ChatGPT, Perplexity and Google AI

This page is the argument; that one is the method. Read it next — it is the direct sequel.

Rank on Google with Reddit comments

The classic-search half of the same play, with the etiquette rules spelled out.

How to choose social listening keywords

Building the keyword list that surfaces the threads worth being in, without the noise.

How to find leads on Reddit

Turning the same monitoring setup into qualified lead opportunities.

Should you pay someone to post on Reddit?

If the citable comment is the asset, whose account it goes out from is the question this page's playbook runs into.

Reddit keyword monitoring

The most-cited domain across the major AI answer engines — and how to watch it.

Hacker News keyword monitoring

Dense, specific, technical threads — the kind of detail AI answers quote.

Quora keyword monitoring

Questions phrased the way people phrase prompts, answered before anyone else gets there.

Sources

Every statistic on this page comes from the published research below and was last checked in August 2026. Search and AI-citation data moves quickly, so treat the figures as a snapshot of 2026 rather than a permanent state of affairs.

Stop optimising for the click. Start being in the answer.

ParseStream monitors Reddit, X (Twitter), LinkedIn, Quora and Hacker News for the questions your buyers ask, and drafts a reply for you to edit and post yourself. The threads AI answers quote are the ones you have to be in.

Start your free trialRead the GEO & AEO guide

3-day free trial · Cancel anytime · No automated posting