A growing share of the internet is built for crawlers, counters and AI answers rather than for readers, and the useful question is who a page was made for, not whether AI wrote it.
- Cheap pages make volume the strategy. When a page costs almost nothing, a tiny return per visit is enough, with no one needing to intend harm.
- Four kinds of room. Filler written to be crawled, shops built from generated photos, descriptions and reviews, audiences made of bots, and pages shaped to be quoted by AI.
- Engagement is no longer evidence. Automated traffic now outweighs human traffic, so a like count shows what someone was willing to pay for, not what people think.
- Writing for AI answers is not spam by default. What gets a page quoted is mostly clear structure, accuracy and visible sourcing; the test is whether the specifics hold up.
- Do not decide everyone is a bot. Most of what you meet online is still made by people; look for a specific, checkable detail that came from experience.
If a page costs almost nothing to produce, and even a tiny fraction of arrivals generate revenue — an ad impression, an affiliate click, a harvested email — then volume becomes the entire strategy. Quality stops being the mechanism by which money is made. That single sentence explains every category below, and it explains them without anyone needing to intend harm.
Four kinds of room
The page written to be crawled
Long, fluent, competent, and containing nothing that required anyone to do the thing it describes. It answers the question you searched — eventually, after several hundred words explaining that the question is important.
Roughly half of newly published articles are now primarily machine-generated, and that share plateaued rather than continuing to climb. The volume is real; the flood did not become a takeover.Graphite, “AI Now Writes as Many Online Articles as Humans”, May 2026, read at source 23 Sep 2026: “since Q1 2025 the percentage of primarily AI-generated articles has plateaued at roughly 50%.” More counts, with dates, in the quarterly record.
- Length without specificity — many words, few facts that could be checked.
- The answer buried under preamble about why the topic matters.
- Your search phrase repeated verbatim as headings.
- No author who exists anywhere else, no date, no correction history.
- The decisive test: is there a single detail here that could only come from having actually done this? A price paid, a thing that went wrong, a name of a specific product that failed.
The shop with nothing behind it
A storefront where the photographs are generated, the descriptions are generated, the reviews are generated, and nobody involved has handled the item. Some are ordinary dropshipping with synthetic packaging. Some exist only to occupy a search result or collect a card number.
- Product photos with impossible reflections, inconsistent shadows, or text on packaging that dissolves under zoom.
- Reviews that praise in general terms and never mention a specific use, flaw or delivery experience.
- A catalogue with no coherent theme — the same shop selling garden tools, skincare and phone cases.
- No physical address, no company registration, and a returns policy that is grammatically perfect and procedurally impossible.
The audience that is not there
Engagement is the cheapest thing on the internet to manufacture — views, likes, replies, and entire conversations between accounts nobody runs. Automated activity accounted for 53% of measured web traffic in 2025, the second consecutive year machines outnumbered people.Imperva/Thales Bad Bot Report, Apr 2026, full-year 2025 telemetry, read at source 9 Sep 2026: “automated traffic continues to outpace human activity online, accounting for more than 53% of all web traffic in 2025, up from 51% the year before”. Vendor measurement, not a universal census.
The consequence is not that your friends are bots. It is that numbers stopped being evidence. A post with fifty thousand likes tells you what somebody was willing to pay for, not what people think.
- Engagement wildly out of proportion to the substance of what was said.
- Replies that agree enthusiastically without engaging with any specific claim.
- Accounts with history but no continuity — opinions that do not connect to each other over time.
- The same phrasing appearing across many accounts within a short window.
The page written to be quoted by a machine
The newest room, and the most interesting, because it is not spam. Search optimisation had a twenty-year head start; its successor optimises for being cited inside an AI answer rather than ranked in a list of links.
It goes by several names — answer engine optimisation (being the extracted answer), generative engine optimisation (being the source an AI trusts and recommends), and a handful of others. Wikipedia standardises on GEO and treats the rest as synonyms; a common counter-argument, including from people in search, is that the whole practice is still just SEO under a new name. There is now a tooling industry around it, selling citation-share monitoring at hundreds of dollars a month.Not read at source: terminology and industry state recorded Aug 2026, but the individual documents were not recorded at the time, so treat this paragraph as reported. An earlier version attributed the “still just SEO” view to Google’s own 2026 documentation without a link; that attribution has been withdrawn.
What makes it different from the rooms above: the tactics that work are largely the tactics of being genuinely useful — clear structure, direct answers, factual accuracy, visible sourcing, being corroborated elsewhere. Machines reward legibility. So a page optimised to be quoted often reads better than a page optimised to be clicked.
- An answer in the first sentence, before any context — written to be extracted cleanly.
- Heavy use of question-shaped headings, comparison tables and bullet summaries.
- Statistics with a named source attached, positioned to be quotable.
- The honest caveat: every item above is also what good writing looks like. This is the one room where the optimisation and the reader's interest genuinely overlap — which is why the tell is not the format, it is whether the specifics survive checking.
This site publishes an llms.txt summarising its own findings for AI systems, uses structured data, front-loads answers, and names sources beside figures. That is generative engine optimisation, done deliberately. The distinction it holds to is narrow but real: everything optimised here is also true and checkable. The format is shaped for machines; the content is not written for them. When those two ever conflict, the machines lose.
In any of the four rooms, look for one detail that could only come from doing the thing. Format alone tells you little, because good writing and machine-shaped writing share it.
What this is not
It is worth being precise, because the strong version of this idea is wrong and corrosive. Most of what you encounter online is still made by people. Human activity remains a large minority of traffic and the majority of what actually gets read. The claim on this page is narrower: a growing share of the internet is addressed to systems rather than to you, and once you can see the difference it is difficult to unsee.
The failure mode to avoid is deciding everyone is a bot. That belief costs you real relationships and gains you nothing — see the theory, and where it overshoots.
Ask who this was made for. Not "is this AI" — that question is often unanswerable and rarely useful. Ask whether the page, the listing or the thread was addressed to a person or to a counter. Empty rooms are built to be walked through by machines, and they almost never contain the one thing that would prove a human was here: a specific, checkable detail that came from experience.
Keep the suspicion narrow: most of what you read is still made by people, so ask who a page was made for rather than assuming the worst about everyone.