Are you in the weights? How AI decides what to recommend

Back in June, a little website called In the Weights took over my feed. You type in a name. It tells you how well the big AI models know that person without searching the web. Founders, investors and podcasters all ran their own names and posted the results. Some were thrilled. Some were quietly crushed.

It was a vanity toy. It also asked the most important question in AI search, and almost nobody noticed.

Last week my friend Geri Ruçi sent me a piece by David McSweeney at QueryBurst on how ChatGPT builds its answers. Read the two together and you get the clearest picture I've seen of how brands win or lose in AI answers. So here's the plain-English version, for anyone who keeps hearing "GEO" and nodding politely.

What "the weights" means

An AI model learns by reading a vast slice of the internet, books and more. It doesn't store those pages. It squeezes what it learned into billions of numbers. Those numbers are the weights.

If you're "in the weights", the model knows you from memory. Ask it about your sector and your name comes to mind unprompted, like a well-read friend recommending a restaurant.

If you're not in the weights, the model has to go and look you up. And it only looks up what it already suspects is worth finding.

The model has an opinion before it searches

This is the bit most people miss. When you ask ChatGPT "which CRM should I buy?", it doesn't start with a blank page. McSweeney's argument is that it drafts an answer first, from the weights. You never see this draft. It happens behind the scenes, in seconds.

The draft already has a shape. It names the brands the model expects to recommend. It lists the points it plans to make. It might not be final, but it sets the agenda for everything that follows.

That's why In the Weights matters. It's a crude window onto that hidden first draft. If a model doesn't know you, you're missing from the draft before the race has started.

Search is a fact-check, not a hunt

Next, the model turns its draft into a research plan. Each claim becomes a search. The industry calls these "fan-outs". For the CRM question they might look like this:

• best CRM for small teams 2026 reviews
• HubSpot vs Pipedrive comparison
• HubSpot pricing plans 2026
• Pipedrive reviews onboarding

Spot the pattern. One broad search, then a run of searches that test the brands already in the draft. The model is mostly checking it was right.

It then pulls back pages, scores the passages against what it expected to find, and drops the rest. The survivors ground the final answer and earn the citation links. Your page can be found and still lose at the scoring stage.

The model can change its mind. A new brand can break in. But it has to get found, survive the cut and make a stronger case than the favourite. That's a much harder job than being in the draft from the start.

How an AI answer gets made

1. The question arrives with context. Your words, plus your job, past chats and saved memories.
2. The model drafts from the weights. A hidden first answer that already names the brands it expects to recommend.
3. The draft becomes a research plan. Each claim turns into a fan-out search.
4. Pages come back. One broad search, then searches that test the brands in the draft.
5. Passages are scored. Each is checked against what the draft expected. The rest are dropped.
6. The survivors ground the answer. They shape the final reply and earn the citation links.

The draft sets the searches, so a brand missing from the draft starts every later step behind.

Who's asking changes everything

ChatGPT knows things about you. Your job, your past chats, what you've told it to remember. All of that travels with every question.

So a plumber in Portsmouth and a marketing director in Manchester can type the same words and get different answers. Different context means a different draft. A different draft means different searches. Different searches mean different winners.

This is why the screenshots people share of "what ChatGPT says about us" only tell part of the story. They show one answer to one person on one day.

Two jobs for every brand

Put it together and GEO comes down to two jobs.

Job one: get into the weights. This is the long game. Models learn from what the web says about you, repeatedly and consistently, across many trusted sources. Press coverage, expert reviews, industry lists, forums, Wikipedia, interviews. Not just your own site. You can't buy your way in next week. You earn it over training cycles, and those cycles are getting shorter.

Job two: win the fact-check. This is the short game. When the model goes looking, your pages need to hold the exact evidence it wants. Clear prices. Plain comparisons. Honest pros and cons. Real reviews on third-party sites. Content written as answers, not as brochures.

Most brands only work on job two, because it looks like SEO. The brands that dominate AI answers do both.

Where to start

1. Ask the model what it knows with search off. Ask ChatGPT, Claude or Gemini about your category without web search. Note who comes up and how you're described. That's your starting position.
2. Find out where you drop out. Were you missing from the first draft? Or found in search and then cut? They need different fixes.
3. Earn mentions off your own site. Get named in the places models trust: trade press, expert roundups, review platforms, community threads.
4. Write pages that answer the checks. Think about the searches the model will run about you. Pricing, comparisons, alternatives, reviews. Make sure the answer is on your page, in plain words.
5. Skip the self-serving listicle. Ranking yourself first in your own "best of" list fools nobody. McSweeney argues it can even reinforce the category leaders.
6. Think in customers, not keywords. Map who's asking and what they've already told the model. A CFO and a junior buyer get different answers.

Are you in the weights? It's worth finding out. If you'd like a hand working out where you stand, get in touch.