What do Google and ChatGPT have in common?

The library was the first time knowledge outlived the knower

Alexandria did not invent writing. It invented the pile. Put enough scrolls in one room and the room knows more than any person in it, and it keeps knowing after they die.

Ancient sources give wildly different counts for how many scrolls were in it, from forty thousand to seven hundred thousand. The classicist Roger Bagnall argued in 2002 that essentially all of those numbers are inflated fantasy. Even the founding myth of information storage is a claim nobody checked.

Google never read those pages

This is the part almost nobody internalised, and the whole essay turns on it. Google did not evaluate whether a page was good. It counted who pointed at it.

Brin and Page said so themselves in the 1998 Stanford paper that described the system. Their words: "Academic citation literature has been applied to the web, largely by counting citations."

The patent went to Stanford, not Google. Number 6,285,999, filed January 1998, granted on the fourth of September 2001. Exactly three years to the day after Google filed for incorporation.

It also taught us to talk like machines

Nobody says best running shoes flat feet 2026 out loud to another human being. We learned to strip our questions down into keyword shrapnel because that is what the box could handle.

In 1998 the average AltaVista search ran 2.35 words. That number comes from a log of 575 million real queries, which makes it one of the better-sourced facts in this entire piece. By 2025 the average Google query had reached 4.0 words.

For twenty-four years, we did the adapting.

The difference was never intelligence

It was that the box talked back.

OpenAI's own launch note described a system that could "answer followup questions, admit its mistakes, challenge incorrect premises, and reject inappropriate requests." Read that list again. Not one item on it is about knowing more things. Every item is about conduct.

They shipped it as a free research preview and expected a quiet week. Five days later Altman posted that it had crossed a million users. By October 2025 OpenAI was reporting 800 million people using it every week, and 900 million by February 2026.

The splash screen offered three example prompts, and they tell you exactly who OpenAI thought would show up. Explain quantum computing in simple terms. Got any creative ideas for a 10 year old's birthday? How do I make an HTTP request in Javascript?

A physicist, a parent and a developer. They had no idea.

III. What they actually share

So what do they have in common?

The lazy answer is that they are both search. People say it constantly, and it is wrong in a way that matters, because it is the reason most advice about AI visibility is worthless.

If the two things were just search with different paint, the playbook would transfer unchanged. It does not transfer. Anyone who has tried already knows that.

Prediction is the mechanism. Consensus is the fuel.

Both systems output the most agreed-upon thing. One did it by counting citations between documents. The other does it by counting co-occurrence across a corpus. The plumbing is different and the logic is identical.

Which means the shape of the job changed underneath everyone.

There is no position ten in a paragraph. There is no page two of an answer. Either the model names you or it does not, and that decision was made long before the person asked, by everything the internet had already said about you.

Then the window priced in

Here is what that cost, in numbers nobody disputes because they came out of SEC filings.

In the third quarter of 1998, the average paid click on GoTo.com cost three cents. By 2001 it was twenty cents. In 2026 the average Google Ads click costs $5.42.

That is roughly twenty-seven times the 2001 price. Inflation over the same period was about 1.8 times. The rest is what it costs to arrive late.

The gap between knowing and everybody knowing lasted about three years. Then it closed, and it has stayed closed for twenty.

We are in the equivalent moment right now, and most people are spending it arguing about whether it is happening.

One correction before you get comfortable

Google is not dying. That story gets repeated constantly and the numbers do not support it. StatCounter had Google at 91.3% of global search in July 2026, up from 89.9% in March. The 2024 dip recovered.

This is not a story about replacement. It is a story about a second machine running the same logic next to the first one, and about the fact that both of them now decide whether you exist.

The way you win an open window has not changed since I built and sold an agency by being the split testing guy. You own one position. Not a keyword list, not a content calendar. One position, stated in one sentence, that you are willing to be boring about for three years.

V. How people actually decide

Step away from marketing for a second

Think about the last thing you bought that cost more than a dinner. You did not build a spreadsheet of ten options and score them on weighted criteria. Nobody has ever done that.

You narrowed to two or three almost immediately, on grounds you would struggle to defend, and then you looked for permission to pick the one you already wanted.

In person, a conversation comes with scaffolding

You never notice it because it is free. The other person asks a follow-up you did not think to ask. You hear them hesitate before the word "probably" and you weight it correctly. You watch their face lose interest and you change tack.

Most of all, you get permission to say "wait, what?" without feeling stupid. That single affordance is doing enormous work, and no interface offered it for twenty-four years.

Chat put the scaffolding back

That is the real reason adoption looked the way it did. Not raw capability. Manners.

Go back to OpenAI's launch list one more time: answer follow-up questions, admit mistakes, challenge incorrect premises. That is a description of a decent conversational partner. It is not a description of a better index.

People did not switch because the machine got smarter. They switched because it finally behaved like the person they had been trying to find since the cave.

What it already cost

Wikipedia is the canary

In October 2025 the Wikimedia Foundation reported that human pageviews had fallen about 8% year over year. They had just rebuilt their bot detection, stripped out traffic that had been evading it, and found real human demand underneath had dropped.

Roughly 90% of Wikipedia's visitors have historically arrived through Google search. The Foundation's own explanation was direct: search engines increasingly answer the question instead of linking to the site that answered it.

Cloudflare put a number on the trade. In July 2025, for every visitor Google sent to a site, it crawled about five pages. OpenAI crawled about 1,091 pages per referral. Anthropic crawled about 38,000.

By mid-2026 Cloudflare reported that more than half of all internet traffic was non-human, and that for every hour people spend looking for information, about fifteen minutes of it happens on the open web.

That is the deal on the table. The machines read everything and send back almost nothing. Being read is now the product.

VII. The playbook · One

Baseline before you touch anything

Write twenty-five prompts a real buyer would type. Not keywords. Sentences, with the mess left in. Run every one through ChatGPT, Perplexity, Google AI Mode and Claude, and log three things: were you named, were you cited, and who got named instead.

That grid is your GA4 for LLMs. Most brands have never once looked at it, which is remarkable given how much they spend guessing.

Do it this week: twenty-five prompts, four engines, one spreadsheet. It takes an afternoon and it will change what you argue about in your next meeting.

Two

Own one position

Pick the single question you intend to be the answer to, and write it as one sentence: when somebody asks X, the answer is us.

I sold an agency on the back of one of these. I was the split testing guy. Not the growth guy, not the CRO guy, not the full-funnel partner. Narrow enough to be boring, specific enough to be repeatable by someone else at a dinner party.

That last part is the whole test. A consensus machine cannot amplify a position that other people cannot restate.

Three

Earn agreement somewhere other than your own website

Your site is one vote. It is a vote the model knows you control, which is exactly why it is worth less than you want it to be.

Ahrefs ran an experiment through mid-2026 across nearly ten thousand AI answers. For a brand that already had third-party coverage, 94% of new brand mentions came from content the brand did not own. Only 6% came from its own pages.

Now the honest complication, because this is where most GEO advice lies to you. For a brand with no third-party coverage at all, the numbers flipped: 82% of mentions came from its own pages. If nobody is talking about you yet, your own content is the only vote there is, and it works.

And the largest citation study anyone has published, 11.84 billion citations analysed by Profound between April and July 2026, found that company websites make up about 57% of all AI citations. Earned media was 23%. Social was 13%.

So "Reddit runs the models" is true at the very top of the distribution and false across the whole of it. Your own site still matters enormously. It is just not sufficient on its own once other people start talking about your category.

Do it this week: pick the three surfaces where your buyers already argue with each other, and go be genuinely useful on them under your own name.

One warning worth taking seriously. Google's spam policies were updated on 15 May 2026 to name "attempting to manipulate generative AI responses in Google Search" as spam. There is a real line between earning coverage and manufacturing it, and the people who spent 2005 buying links already know exactly where it is.

Five

Write to be lifted

There is peer-reviewed evidence for this, which is rarer in this field than the confidence levels suggest. The KDD 2024 paper that named generative engine optimization tested nine ways of editing a page and measured visibility across ten thousand queries.

Adding quotations lifted visibility 42.6%. Adding statistics lifted it 32.8%. Citing sources lifted it 27.7%. Keyword stuffing lowered it 8.8%, which should settle an argument some people are still having.

Do it this week: put a self-contained summary above your intro that answers the question completely on its own, turn your headings into the questions people actually ask, and add a comparison table to anything with commercial intent. Models lift tables constantly.

Six

Every claim gets a number, a date, or a name

"Studies show" is uncitable. A model cannot pass it along, because there is nothing to pass. "According to the Ahrefs difference-in-differences study published in May 2026" is a sentence a machine can hand to somebody else without losing anything.

This is the single highest-leverage editing rule I know, and the statistics finding above is the evidence for it. It is also why this page carries a source line under almost every section, including the ones where the source undercuts me.

Do it this week: open your best-performing page and count the unsourced claims. Fix the top five.

Seven

Entity hygiene, and I do mean hygiene

Point your organisation markup at Wikipedia, Wikidata and your verified profiles. Keep one consistent identifier across your graph. Make your brand facts identical everywhere they appear, because a model resolving you to two entities is a model that trusts neither.

This is disambiguation. It is not a growth lever, and anybody selling it as one is about to become the subject of the next section.

The AI Mode and ChatGPT effects came in at plus 2.4% and plus 2.2%, neither significant. Four separate statistical tests agreed. The study found no vertical, no authority level and no schema type where it helped.

A second study, published on SSRN in February 2026, found something sharper. Pages with attribute-rich schema were cited 61.7% of the time. Pages with no schema at all were cited 59.8%. The difference between them was not statistically significant, at p equals 0.71.

Generic schema, the Article and Organization markup most sites actually ship, scored 41.6%. Worse than nothing.

The strongest argument against me

I am not going to pretend this is settled. In April 2026, AirOps and Kevin Indig analysed 16,851 queries across 353,799 pages and found pages with JSON-LD were cited 38.5% of the time versus 32.0% without. They controlled for word count, heading count and domain authority, and called it an independent signal.

Here is why I still think they are measuring a shadow. They did not control for retrieval rank, which was their own single biggest finding. Position one was cited 58.4% of the time. Position ten, 14.2%. That effect is four times larger than the schema gap they found, and pages that bother with structured data are systematically the pages that rank.

The Ahrefs design holds the page constant and measures before and after. That is a stronger instrument for a causal question. Read both and decide for yourself, which is more than most people in this argument will offer you.

What to do: ship schema. Ship it for entity hygiene and rich results, which are real. Do not ship it expecting citations, and do not let anyone bill you for it as an AI visibility service.

Two and three

AI did not kill SEO, and llms.txt is not the new robots.txt

Take the second one first. Ahrefs looked at 137,210 domains in mid-2026. About 28% had published an llms.txt file. Of those, 97% received no requests at all in the month studied. Not from AI crawlers, not from anything.

Google's documentation was updated in June 2026 to say plainly that these files are not needed for Search. Asked directly whether Google publishing its own llms.txt was an endorsement, John Mueller answered: "but to be direct, no."

OpenAI, Anthropic and Perplexity all publish an llms.txt for their own developer docs. None of them documents its crawler reading yours. People keep mistaking the first thing for the second.

Publishing the file costs you nothing. This page has one. Believing it does something is the expensive part.

On the bigger myth: search rank is still the strongest surviving predictor of whether a model cites you. The SSRN study found rank position dominated every other variable at p below 0.001, with position one cited in 43% of queries and position seven in 5%. The AirOps data says the same thing from the other direction.

GEO is SEO plus reputation. It is not instead of it. Anybody telling you to stop doing search work is selling you the replacement.

What did change is what a click is worth. Pew found that when an AI summary appears, 8% of people click a traditional result, against 15% when it does not. Ahrefs measured a 58% drop in first-position click-through by December 2025. And 68% of US Google searches now end without a click at all.

Fewer clicks, and the ones that survive are worth more. Semrush put the average AI search visitor at 4.4 times the value of an organic one, measured by conversion rate.

Lily Ray tested 100 commercial queries across AI Overviews in mid-2026. When a brand's own "best of" page got cited, that brand was left out of the actual recommendation 69% of the time. The model read their page, took the list, and recommended a competitor from it.

The Ahrefs experiment found the same pattern independently, at 43%. Two different teams, two methods, same uncomfortable answer.

So a dashboard showing your citations climbing can be showing you a machine that is politely using your research to sell somebody else's product. Measure whether you are named in the answer, not whether you appear in the footnotes.