A client sends me his editorial plan for the half-year. Twenty-four topics. I read the titles: "what is lead nurturing", "the 7 benefits of a CRM", "how to choose invoicing software". I answer in one sentence: twenty-one of those, ChatGPT already answers better than you will, and for free. The three that remain are worth the whole half-year.
This article says which three, and why. Not a prophecy about the death of SEO, a sorting exercise. What changed is not the usefulness of content, it is the value of the middleman. An article that neatly restated what already existed elsewhere had value as long as the reader had to go through a page to get it. It no longer does.
Generative AI did not kill content. It killed the content that repeated what already existed elsewhere. That category is far broader than most people are willing to admit.
What the numbers actually say
Let us start with facts, because this subject attracts a lot of opinion and very little measurement. The Pew Research Center tracked the real behaviour of more than 900 users across nearly 69,000 Google queries: when an AI summary appears, clicks on a classic result fall to 8% of visits, against 15% without a summary. Clicks on a link inside the summary account for just 1% of visits. On pages with a summary, 26% of sessions end there, against 16% without.
Ahrefs, for its part, compared 300,000 keywords before and after the AI Overviews rollout: a 34.5% lower click-through rate in first position when an AI summary is present. The detail that matters sits in the methodology: 99.2% of the queries triggering those summaries are informational.
Hold on to that last figure, it is the most useful one in this article. It is not "traffic" that collapses. It is one page type. Your service pages, product pages, pricing pages and local pages are barely affected. Your twenty-four definition articles are.
Generic content is not threatened, it has already been replaced
I say it plainly in meetings, even when it makes people wince: stop producing those pages. Not optimising them, producing them. A "what is X" article only ever had one quality, being a decent middleman between a simple question and a known answer. The machine is a better middleman. It is faster, it does not throw a newsletter pop-up, and it answers in the reader's language.
One nuance, because I am not dogmatic here. Those pages keep an internal function: they structure a topic cluster, they carry internal linking and they establish topical authority. But one well-built pillar page fills that role. Fifteen variations on the same definition do not fill it better, they dilute it.
The three raw materials the machine does not have
A language model works on what has been published. It therefore cannot produce what has never been published. That is where everything you have left sits, and there is more of it than people think.
1. First-hand data
Your numbers. The return rate per catalogue category, the average delay between first contact and signature, the real acquisition cost in your sector. On an event ticketing account I ran on Meta Ads, the cost per ticket sat around 2.66 €; for a training provider on lead generation, customer acquisition cost came in at 3.73 €. Neither figure exists anywhere except in the accounts I managed. No model invents them, and when they are published, they get cited.
2. A stated position
An AI averages the available opinions. It takes no risk, because it has none to take. You do. "On this market I advise against 300 € links, coherent placements at 30-50 € are enough": that sentence commits someone. It is useful precisely because it can be wrong. An editorial plan without a single stated position is an editorial plan the machine will write better than you.
3. Dated, situated experience
Not "local search best practices", but what happened on a specific account, on a specific date, with the result obtained and what did not work. Google asks for exactly that in its official documentation on creating helpful, reliable, people-first content: show first-hand experience, and be able to answer three questions, who wrote it, how it was produced, and why. That framework is called E-E-A-T (experience, expertise, authoritativeness, trustworthiness). It is not a technical criterion, it is a demand for proof.
Five content types, and what is left of them
The table below is the one I use to sort an existing stock of articles. Read it top to bottom: the first two rows are to be cut back, the last three to be developed.
| Content type | What AI does with it | What you keep | What to do |
|---|---|---|---|
| Definition, "what is" | Summarised in full inside the answer | Internal linking, nothing more | One pillar page, not fifteen variations |
| Generic tutorial | Copied step by step | Almost nothing | Redo it with your screenshots, your settings, your failures |
| Untested tool comparison | Synthesised from other comparisons | Nothing | Actually test, publish the method and the numbers |
| Study or proprietary data | Cited, with the source | The citation, the brand, the inbound links | Produce some, even at small scale |
| Product, pricing, free tool page | Sent as a destination | The click and the conversion | Work on clarity, not word count |
The last row deserves a word. Ahrefs published an observation on its own traffic: 80% of visits coming from AI engines land on the homepage, product pages and free tools, not on blog articles. The machine summarises the knowledge, then it sends people to a destination. So the destination has to exist and be readable.
Classic SEO is not dead, it has become the toll gate
This is the most widespread misreading right now, and an expensive one. Ahrefs analysed 1.9 million citations from one million AI Overviews: 76% of those citations come from pages already ranking in the organic top 10. In other words, you do not choose between SEO and GEO (being cited by ChatGPT, Perplexity or Google's AI answers). The second is largely won with the first. A slow, poorly indexed site with no topical authority will not be cited any more than it is ranked. Technical work and generative engine optimisation pull in the same direction.
The most expensive mistake
Watching traffic fall and publishing more. It is the reflex I meet most often, and it is the worst one. Ahrefs surveyed 879 marketers: those using AI publish 42% more content. Doubling the output of content the machine already summarises means doubling the quantity of lead. The right move is the opposite: produce less, produce what only you can write, and move the budget towards data and destination pages.
How I rewrite an editorial plan
Whatever the client, I work in the same order every time. Less spectacular than a redesign, and far more profitable.
- Sort the stock into three piles. Keep as is, merge (three weak articles on the same subject become one pillar page), delete and redirect. On a typical company blog, the "merge" pile is often half the volume.
- Find the data you already own without knowing it. The back office, support tickets, ad accounts, customer reviews. There is almost always enough for two or three data-backed pieces nobody else can write.
- One stated position per quarter. A subject on which the company decides, argues and commits. That is what gets relayed, cited and debated.
- Point the internal links at the destinations. If AI engines send people to product and service pages, articles may as well lead there explicitly, with descriptive anchors.
Measure differently, or you will give up too early
Last point, and a decisive one. If you keep judging an editorial plan by the traffic of its informational articles, you will conclude it failed whatever happens, since that traffic falls mechanically. So change instrument: how often the brand is cited in AI answers (tracked through recurring prompts, as I detail in my piece on GEO monitoring), branded search volume, and assisted conversions.
Good news on that front: Semrush estimates that a visitor arriving from an AI search is worth roughly 4.4 times a classic organic visitor in conversion rate, because they arrive having already compared. Volume falls, quality rises. One caution, and a serious one: the same house tracked 230,000 prompts and more than 100 million citations over thirteen weeks, and watched Reddit's citation share on ChatGPT drop from around 60% to 10% in a month. Never build a strategy on the citation habits of a single platform. They change without notice, and without explanation.
In the crucible: what remains once the machine has summarised everything
The lead, here, is the content that repeats. It filled years of editorial calendars and plenty of output reports, and its value collapsed in under two years. The gold is what only you can write: your numbers, your judgement calls, your documented failures. That material cannot be synthesised, because it exists nowhere else to be synthesised from.
At bottom, generative AI does us an uncomfortable favour: it makes mediocre content permanently unprofitable. You can live it as a loss. I see it rather as a return to the only question worth asking, the one we should have asked earlier: what does this text give the reader that they cannot find elsewhere? I take that question on site by site as part of my content strategy work.
Does your editorial plan still hold up?
I sort through your existing articles, identify the data you already own but do not use, and hand back a tighter plan: fewer pages, more reasons to be cited.