Google isn’t the front door anymore. Your blog still has a job.

In February 2024 Gartner predicted a 25% drop in traditional search engine volume by 2026, and every content team I know quoted it in a planning deck at least once, mine included. 2026 arrived, and Graphite’s analysis with Similarweb of the top 40,000+ US sites found organic search traffic down about 2.5% year on year, comparing February to December 2024 with January to November 2025. The ten largest sites gained 1.6% on average, and the losses bunched among sites ranked roughly 100 to 10,000. Graphite’s CEO Ethan Smith concluded “the prevailing narrative that SEO traffic has dramatically declined is false”, and he says in the same piece that he has a financial interest in that conclusion. So the organic traffic decline everyone braced for didn’t show up in the aggregate, and anyone whose blog got flattened this year knows that’s a useless conclusion, because the aggregate was never the figure that applied to them.

Where is the organic traffic decline concentrated?

The organic traffic decline is concentrated in queries a paragraph can close, which is why publishers lost far more than the web as a whole. Chartbeat data across more than 2,500 news sites, published in the Reuters Institute’s 2026 trends report, showed Google organic traffic down 33% globally and 38% in the US between November 2024 and November 2025. The 280 media leaders in the same report expect search traffic to fall about 43% over the next three years, and around one in five expect to lose more than 75%.

Dataset	Population	What it found
Graphite and Similarweb, January 2026	Top 40,000+ US sites.	Organic down about 2.5%; the top 10 up 1.6%; sites ranked about 100 to 10,000 lost most.
Chartbeat via Reuters Institute, January 2026	2,500+ news sites.	Google organic down 33% globally and 38% in the US.
Pew Research Center, July 2025	900 US adults, 68,879 searches.	Clicks on a traditional result fell from 15% to 8% when an AI summary appeared.

These datasets measure different populations, and what separates them is less industry or site size than query type. Traffic from questions a summary can answer has gone, while queries where the searcher still needs the page held up. So the organic traffic decline reads as a rounding error or a redundancy depending on who’s describing it, and “is organic traffic down?” has no useful answer at that level. Ask which queries you lost, and whether you were ever making money from them.

Diagnose before you blame the summary, too. When I audited this site’s first six months in Search Console, roughly 50 articles produced 828 impressions and 14 clicks, and the cause turned out to be position on page two, which no amount of worrying about AI Overviews would have fixed.

What is the 67-word test?

The 67-word test checks whether a page survives AI summarisation by asking whether its headline can be fully answered in 67 words. The figure comes from the Pew Research Center, which tracked 900 US adults across 68,879 Google searches in March 2025, still the cleanest behavioural dataset on what AI summaries do to clicks. Users clicked a traditional result 8% of the time when a summary appeared, against 15% when it didn’t, clicked a link inside the summary 1% of the time, and ended their browsing session on 26% of pages with a summary, against 16% without.

The figure I keep coming back to from that study is smaller and more useful: the median AI summary was 67 words long.

Take the page that used to be your best performer and try to write a complete answer to its headline in 67 words: if you can, Google can too, and it already has.

It takes about four minutes a page and tells you more than any audit score, because it goes straight at whether the page can be replaced by a paragraph. A page passes when compression costs the reader what they came for: a dataset nobody else has, a decision that cost somebody money, a comparison that ends in a recommendation, or a method detailed enough that following the 67-word version would get someone fired.

Which content types survive the organic traffic decline?

The content types that survive the organic traffic decline are the ones a summary can’t compress without losing what the reader came for, and funnel stage has little to do with it: plenty of decision-stage pages got eaten, and plenty of top-of-funnel material held up.

Content type	What an AI answer does to it	Verdict	What to do
Definitional explainers (“what is X”)	Absorbs it completely; the summary is the article.	Dead as a traffic play.	Keep one, as a definition source for citation, and stop producing more.
Generic how-to guides	Lifts and reorders the steps; no reason to click.	Mostly dead.	Rewrite around a specific implementation with the figures and the mistakes.
Original research and proprietary data	Cites it heavily, and the citation names you.	Strongest surviving format.	Produce more of it; this is where the budget should move.
Opinion with a stake	Can summarise it, can’t replace it, because people want the person who holds it.	Resilient.	Take positions someone can argue with.
Comparisons that end in a recommendation	Pulls them into shortlist answers constantly.	Resilient and commercially loaded.	Include the cases where you lose; that’s what makes buyers trust the rest.
Product, pricing, security and integration docs	Barely touches them; the buyer still has to come and read the page.	Untouched, for now.	Treat as pipeline infrastructure and give it an owner.

The two rows I’d fight for are original research and the comparison. I’ve made the case for the first in original research as content strategy, and for the second in the B2B comparison page. Comparisons are also why the technical pages I built for CIO and CTO buyers at a B2B SaaS company serving 200+ higher education institutions kept converting while the awareness-stage library got quieter, a build I wrote up in technical content for CIOs. The documentation row has its own piece, on why product documentation is outranking the blog in AI answers.

What is a blog for when buyers start in AI?

A blog’s job now is to supply the answers buyers use to build their shortlist, and that job pays off whether or not the reader ever visits. G2 surveyed 1,076 B2B software buyers in March 2026 and found 51% now start their research with an AI chatbot more often than with Google, up from 29% a year earlier. More striking, 69% chose a different vendor than they’d planned to based on the chatbot’s guidance, and 33% bought from a vendor they weren’t familiar with. G2’s Tim Sanders described it as “the third compression era of the buyer journey”. Semrush’s survey of 519 US professionals who use AI at work found 72% use it during early research and 61% compare vendors directly inside the tool.

Read those together and the blog’s job moves upstream of the visit: it supplies the answer that decides whether a vendor becomes a candidate at all, and a third of the time that’s a company the buyer couldn’t have named an hour earlier. An article that gets cited in a comparison answer and produces zero sessions does more for pipeline than a listicle that pulled 400 visitors with no buying intent behind them.

Getting cited is a separate discipline with its own rules, and I’ve documented mine in the AEO/LLMO content optimisation framework. The short version is that substance earns the citation and structure only delivers it, so a green audit score on a thin page still won’t get you cited.

What should you report when sessions fall and nothing is wrong?

Report citation share first, then AI referral quality, then branded search and self-reported attribution, and keep traffic as a diagnostic line underneath. If your monthly report leads with traffic, the organic traffic decline will read as failure every month until someone kills the budget, so change what leads and say why: the measurement stopped describing the outcome some time in 2025.

• Citation share leads. Run a fixed set of the prompts your personas ask, monthly, across ChatGPT, Claude, Perplexity and Gemini, and log whether you appear, who appears instead, and what the answer says about your category. I’ve written up how I run this in AI citation tracking, and where the prompts come from in prompt research.
• AI referral quality comes next, because the volume is tiny and behaves differently from standard organic. The one multi-site study that tested significance found B2B AI referrals converting at close to double organic, while the overall difference disappeared. Segmenting them in GA4 takes about twenty minutes, and the setup is in track AI traffic in GA4.
• Branded search and self-reported attribution close it out. Rising branded queries against falling non-branded traffic is what it looks like when a brand gets discovered somewhere your analytics can’t see. A free-text “how did you hear about us” field on the form is easy to add, and in my experience it tells a different story from the CRM; the full setup is in measuring AI visibility without analytics.

Traffic stays in the report as a diagnostic. At the B2B SaaS company where I tracked 17+ KPIs a quarter, the two results that got quoted back to me in leadership meetings were content contributing to 48% marketing-sourced pipeline in peak months and influencing €230K+ in pipeline in a single reporting period, and neither needed a sessions chart to make sense. I’ve laid out the full hierarchy in content marketing metrics.

Where does this argument run out?

I’m reasoning from a little over two years of data since AI Overviews launched in the US in May 2024 and a handful of libraries, which isn’t enough to be certain about anything.

Citation tracking is still a weak measurement discipline. SparkToro’s Rand Fishkin and Gumshoe’s Patrick O’Donnell ran 2,961 prompts across ChatGPT, Claude and Google’s AI and found less than a 1 in 100 chance that ChatGPT or Google’s AI returned the same list of brands for the same prompt twice. Fishkin’s verdict on tools selling an AI rank was that they’re “full of baloney”. The same research found a brand’s visibility across many runs steadier than its position, which is why I track share of appearances and ignore rank, and why a fixed prompt set run monthly beats a one-off check.

I’d also push back on my own table: “Untouched” is doing optimistic work in that last row, and agentic buying assistants that read pricing and security documentation on the buyer’s behalf will change it. My bet is that those pages get more valuable while the session data around them gets stranger, and I’ve thought through what that does to content structure in content for AI agents.

What I’d do differently: I spent 2024 optimising for the click, and I should have started measuring citations six months before I did.

Working out whether your blog was a traffic channel and nothing else, or an argument with receipts attached, is worth an afternoon with a spreadsheet and the 67-word test, which is roughly where I’d start if I walked into your library cold. If it was only ever a traffic channel, tell whoever funds it that the channel is gone. The follow-on question, whether the blog is a blog at all anymore, I’ve argued out in stop running a blog, start running a media property.