How AI search is killing comparison sites — and what gets more valuable instead

The content middleman is dying, and I say that as someone who has spent years being one.

For most of the last decade, a particular kind of website printed money: take a question people type into Google — “best invoicing tool,” “cheapest business loan in Estonia,” “VAT software for freelancers” — assemble a tidy comparison from information that was already public, rank, and collect a commission when someone clicked through and bought. I have built that page many times. It worked because there was friction between a person’s question and a good answer, and the comparison site sold itself as the thing that removed the friction.

AI search removes that friction directly. Ask an AI Overview, ChatGPT or Perplexity for the “best X for Y,” and you get a synthesised answer — often with a small table, sometimes with caveats — without ever visiting the page that used to sit in that slot. The click that paid for the whole model does not happen. I’m not forecasting this from the sidelines — I see it in my own analytics. That’s one operator’s portfolio, not industry-wide proof; but the broader market points the same way (by various 2026 estimates, a large and rising share of searches end without a click, and AI Overviews measurably reduce through-clicks where they appear — verify current figures, they move).

I do not think this is a doom story, though. I think it is a reallocation. The value does not vanish; it moves. The job now is to understand exactly which layer is being eaten and which layer is quietly getting more valuable — and then to move your weight onto the second one before the first one collapses under you.

What the classic comparison site actually did

Strip away the design and the affiliate disclosures and a classic comparison page does one thing: it re-lists what is already findable. It collects specs, prices and features that the vendors themselves publish, arranges them into a table, adds a paragraph of generically-positive prose per option, and wraps the whole thing in a “best of 2026” headline. The reader gets a shortcut. The publisher gets a click.

The uncomfortable truth is that this layer was always thin. It added convenience, not knowledge. The information existed; the page just saved you the tab-switching. That convenience had real value when search was a list of ten blue links and synthesis was your job. It has almost no value when synthesis is done for you, above the fold, by the engine itself.

AI eats exactly this layer because aggregating and summarising public information is one of the things large language models do best. It’s hard to out-aggregate a tool built to synthesise sources. If your moat was “I gathered the specs into a table,” the moat is gone. The model gathered the same specs, and it gathered them this morning.

The thesis: the content layer commoditises, the data layer appreciates

Here is the reframe I have organised my own portfolio around, and it fits inside the same logic I used when I wrote about the mathematics of a solo business — revenue per durable asset, not revenue per published page.

The content layer is commoditising. Words that describe, summarise or re-list publicly-available facts are now produced at zero marginal cost by the same engines your readers are asking. There is no defensible position in being a slightly-better paragraph about a product the AI can already describe.

The data layer is appreciating. Structured, real, current, first-hand data — numbers the model cannot derive from its training set because they did not exist until you measured them — gets more valuable, not less. Why? Because the AI still has to ground its answer in something. When it says “Tool A processes payouts in two days and Tool B in five,” it needs a source for that claim. If you are the one who actually tested both and published the result, you become the thing it cites. Citation is becoming the new ranking, and you cannot cite a vibe.

So the strategic move is not “write better comparison content.” It is “stop selling the summary and start owning the underlying data the summary is made of.”

A three-zone survival model

Not every niche is affected the same way. I sort the products I cover into three zones, and each zone has a different survival strategy.

Zone 1 — standardised products: become the data source AI cites. When the products are well-specified and comparable on hard numbers (hosting plans, payment processors, payout speeds, fee structures), the winning move is to generate data the vendors do not publish and competitors do not have. Run your own benchmarks. Actually open the account and time the KYC. Measure the real all-in cost on a realistic transaction, not the headline rate. A page that says “I tested seven processors on the same €1,000 EU cross-border sale and here is what each one actually cost” is not a summary an AI can replace — it is a primary source an AI has to reference.

Zone 2 — low-specification needs: build advisory tools on your own data. Many decisions are not “which product is best” but “which is best for me.” That is a computation, not a list. A calculator that takes a reader’s revenue, country and customer mix and returns a ranked recommendation does something an AI Overview cannot do well, because the answer depends on inputs only the reader has. The tool is the content, and your data is the engine underneath it. This is the most defensible position of the three, because it converts your data into an interactive answer instead of a static one.

Zone 3 — low-standardisation: brand authority and first-hand experience. Where products are messy, subjective or trust-laden — and where getting it wrong costs the reader real money — what wins is genuine lived experience and a name people seek out on purpose. An AI can fabricate a plausible-sounding paragraph about “what it is like to run a one-person business in France.” It cannot fabricate the fact that you actually did it, hit the specific wall, and wrote down what happened. Brand is the asset AI cannot mint, because brand is the reason someone types your name instead of a query.

What this means for me, specifically, right now

I am not theorising at you from the sidelines. I run comparison and affiliate sites, and I can watch the AI-driven click erosion in real numbers. Here is what I am actually doing about it across the portfolio.

Generating my own data instead of re-listing vendors’. On the review side of my sites, I am replacing “here are the published features” with “here is what I found when I tested it.” That is slower and it does not scale the way thin content scaled — and that is exactly the point. The cost is the moat. Anything cheap to produce is now produced for free by the machine.

Shipping tools, not just pages. A static “best X” article is the most replaceable thing I own. A calculator that turns a reader’s situation into a personalised shortlist is the least. So the build roadmap is shifting from publishing to tooling — which, conveniently, is the kind of small, sharp, single-purpose software a solo operator can actually ship. It is the same instinct I described in projectologist vs founder: bet on assets that compound, kill the ones the market is making structurally harder.

Leaning into the EU-specific angle a global model underweights. A model trained on the whole internet is, by construction, weakest on local nuance: Estonian e-Residency mechanics, country-by-country VAT-OSS thresholds, the real cost of a cross-border payout inside the single market. A globally-averaged AI gives globally-averaged answers here, and they are often subtly wrong for a specific EU country. That gap is a gift to a publisher willing to go narrow and precise where the model goes broad and vague.

Treating brand as infrastructure, not vanity. Direct, name-led traffic is the one channel AI search does not intermediate. If a reader has decided I am worth coming back to, the Overview box does not stand between us. So the long game is earning that — by being right, being first-hand, and being specific — rather than renting attention from a ranking that is being absorbed into an answer box.

None of this is free. It is more expensive content, more software to maintain, more expertise to genuinely have. But that is the trade the moment is offering: the cheap path is closing, and the expensive path is the one with a moat at the end of it.

The takeaway

Stop competing for the click. Start owning the data and the trust.

The comparison middleman that re-lists findable facts is being disintermediated, and no amount of better prose saves it, because prose is the commodity now. What appreciates is the stuff underneath the prose: data you generated, tools that apply it, and a brand readers choose on purpose. AI search did not kill the value in helping people decide. It just moved that value one layer down — from summarising the answer to being the source the answer is built on.

I would rather be the source than the summary. So that is the layer I am moving to, while the old one still pays for the move.

The flip side of AI eating your clicks is that AI can also send them — being the product the model names is the new ranking. That’s its own discipline: Generative Engine Optimization (GEO), with the real numbers from a product I got AI to recommend from zero.

Are you still publishing summaries, or have you started owning data? If you run an affiliate or comparison site, which of the three zones do most of your products fall into — and what is your honest plan for the AI-Overview slot that used to be yours?