GEO by engine: how AIs arbitrate your sources

A client calls me: he lost 37% of AI clicks

A client calls me on a Tuesday morning. He runs an e-commerce site for spare parts. 843 references. 1,200 indexed pages. He publishes three pieces of content per week and has for eighteen months. He tells me: « The AIs no longer cite me. »

I look at his stats. Classic organic traffic has barely moved. Clicks from ChatGPT dropped 37% in six months. Claude still cites him on generic questions. Gemini ignores him. Perplexity uses him for niche product pages, but not for guides.

For him, everything comes down to arbitration, not content.

Neil Patel published the most complete synthesis to date on this decision layer. He dissects four engines: ChatGPT, Claude, Gemini, Perplexity. He shows how each selects, ranks, then discards competing sources. The logic is not uniform. Each engine has its biases.

The client believed GEO = SEO. He applied the same recipes: keywords, length, internal linking. Result: dense pages, well written, but too slow to deliver the answer. The AIs discard them before line 2.

The decision layer looks at immediate utility. Editorial quality, at this stage, doesn’t weigh much.

I will show you the mechanics. Not theory. The real filters, the signals, the tipping points.

The first client call is a live audit. We look at the page, type the query into ChatGPT, and observe. The diagnosis falls in two minutes. The page is either cited or discarded.

The most striking? The best-written pages are often the least cited. They tell a story. The AI expects an answer. The mismatch costs dearly.

A catalog of 843 references with siloed architecture prevents the AIs from understanding the internal linking. Without semantic linking, each page stands alone. An isolated source loses consensus.

Neil Patel’s data confirms this mismatch. The decision plays out on blocks, not pages. A block = a question + an answer. If the block is drowned, it disappears.

The client wanted to enrich. That was a false trail. It needed to be cut up.

The rest shows you how the four engines weight the filters. And why a response formatted for immediate reward changes the game.

Avant de citer une page, une IA la fait passer par trois filtres successifs. Voici le parcours exact, de la requête à la citation.

The decision layer: three filters before citation

An AI does not read a page like a human. It reads an index, then a prompt, then a list of candidates. Between them, three filters sort.

• Filter 1: selection. The engine pulls 8 to 20 sources depending on the query. Not 50. It cuts fast.
• Filter 2: ranking. It ranks sources by relevance, consensus, freshness, domain authority.
• Filter 3: exclusion. It removes pages that repeat, slow down, or do not answer within the first 300 characters.

What Neil Patel calls the « decision layer » groups these three filters. Each engine weights them differently. That’s why a page can be cited by Claude and ignored by Gemini.

The client at the start thought all engines worked like Google. Wrong. Google ranks pages. AIs select fragments. Granularity changes everything.

A fragment is won on a response block, not on an entire page. If your introduction takes ten seconds to set context, you lose. If it gives the answer in one sentence, you win.

The signals that matter, according to 47 e-commerce deployments:

• Direct answer in the first <h2> or first paragraph.
• Structure in short, tagged blocks, with lists.
• Source consensus: being cited by other pages increases trust.
• Freshness: a visible update date, a recent modification.
• Contextual authority: a site specialized on the semantic field.

Long content remains useful if it is cut up. A 2,000-word page with 18 autonomous blocks is better than a 3,000-word page in continuous flow.

The exclusion filter is the cruelest. It removes a page in 2 seconds. No second chance. It’s the dopamine of the answer: the AI wants immediate informational reward for the user.

If a block does not bring that reward, it goes.

E-commerce product pages often lose here. They start with a brand sentence, then an ambiance sentence, then a spec. The AI, it looks for the dimension, the price, the compatibility. It does not find them in the first 300 characters. It moves to the competitor.

Semantic linking weighs in the selection filter. An isolated page, without parent page, without sister page, is less credible for the AI. It lacks internal consensus.

On a site of 1,200 pages, the selection filter is brutal. The AI cannot read everything. It keeps pages easy to understand. Site structure becomes a signal.

Structured data, <h2> tags, lists: they are shortcuts for the ranking filter. They accelerate extraction. The faster the extraction, the more the source is kept.

The lesson: you don’t win on length. You win on the ability to be extracted in 2 seconds.

ChatGPT, Claude, Gemini, Perplexity: four different arbitrations

The decision layer changes from one engine to another. Neil Patel detailed it in his synthesis. I crossed his tests with my observations on 47 deployments. The result is in the table.

ChatGPT is the most sensitive to consensus. If five competitors repeat the same information, ChatGPT tends to follow them, even if your page is deeper. Freshness also plays: a recent update date improves chances of being cited.

Claude reads more slowly. It appreciates articles structured with clear subheadings, lists, tables. It rewards explainability. A page that explains the reasoning behind the answer has more chances.

Gemini relies on its ecosystem. It crosses with Google Search, Discover, business listings, structured data. A site well anchored in the Google ecosystem has an advantage. But contextual authority prevails: a site specialized on a semantic field beats a generalist.

Perplexity is the most demanding on recency and signature. It prefers signed, dated pages, with verifiable references. Anonymous or undated pages are discarded faster.

The common point: all four engines discard pages that don’t give the answer at the top of the page. That’s the first common filter. Differences come after, on the weighting of secondary signals.

For an e-commerce, that means: you need to win on all four. Not just on ChatGPT. A page formatted for Claude can be invisible to Perplexity if it’s not signed and dated.

The client at the start was cited by Claude but ignored by Gemini. Why? His guides were very well structured, but his site lacked structured data and linked entities. Gemini did not understand the specialization of the domain. He had to strengthen semantic linking and markup.

The decision layer counts four algorithms, not one. Your content must be readable by all four.

DOSE: immediate reward, the real citation lever

The DOSE framework I apply comes from Guillaume Attias / BMO Academy. The letter D, dopamine, is the citation mechanics of AIs. An answer is kept if it brings immediate informational reward.

The AI does not reason like a human. It imitates user satisfaction. If it produces an answer that satisfies in 2 seconds, the user stays. If the user stays, the AI reinforces that selection pattern. Sources that bring that satisfaction win.

Formatting each block for instant utility increases your chances of citation. Concretely:

• The first paragraph gives the answer, not the context.
• Each <h2> asks a precise question.
• Each answer holds in 2 sentences, then deepens.
• Lists and tables are preferred for numeric data.
• The update date is visible.

This formatting is not a hack. It’s the inverted pyramid applied to blocks.

An example. Product page: instead of « Since 2004, our brand imagines sustainable solutions for your workshop », write « The O-ring 47 x 2 mm fits Yamaha 2-stroke engines from 2001 to 2015. Price: 2,40 €. » The first sentence is ambiance. The second is a reward.

The AI keeps the second.

Content that tells a story in the opening is the most discarded. Not because it’s bad. Because it delays the reward.

Across 47 e-commerce deployments, the same restructuring produces the same effect: pages whose first block gives the answer gain citations in 3 to 6 weeks. Pages that keep a narrative introduction stay invisible.

The DOSE framework goes further. Dopamine pushes to create successive micro-rewards: a short answer, a proof, a number, a link to a sister page. Each block must close a loop.

For e-merchants, that’s good news. You don’t need to produce more. You need to reorganize what exists.

The page that wins gives the answer before the user reformulates. AI engines reward that.

Attention: immediate utility does not mean poor content. A short block can be followed by details, specifications, comparisons. But the order changes everything. Answer first, proof next, context last.

Chez ce client pièces détachées, la restructuration fait passer les citations de 9 à 31 sans une seule nouvelle page.

What restructuring yields: the case of 47 deployments

Across 47 e-commerce deployments, the same pattern repeats. Sites that restructure their pages according to the decision layer recover citations in 3 to 6 weeks. Those that publish more without restructuring stagnate.

A spare parts client, 843 references. Before: 9 citations on 50 tracked queries. After: 31 citations on the same 50 queries, in six weeks. Without a single new page. Without ads. Just a rework of semantic architecture and block formatting.

These numbers, I pull from my clients. No published study behind. The order of magnitude: progression goes from 9 to 31 citations depending on sector, existing linking, and competition.

Why does it work? Because restructuring touches all three filters at once.

• Selection: semantic linking connects pages together, and the site becomes coherent.
• Ranking: immediate response blocks send the right signals.
• Exclusion: pages stop wasting time on ambiance sentences.

The first client case from the beginning lost 37% of AI clicks in six months. After restructuring, ChatGPT clicks went back up 22% in eight weeks. Claude stayed stable. Gemini started citing 14 guides, up from 0 before. Perplexity added 8 product pages.

These figures come from citation log tracking. Nothing extrapolated.

No need to rewrite the whole site. We start with the 20 pages that bring the most. They become immediate blocks. Then we extend.

For an e-commerce of 800 to 1,200 pages, the initial effort is 3 weeks. Not 6 months.

Behind citations, there is a working architecture. AIs find pages without friction. Classic SEO follows: click-through rate improves, average cart too.

Another client, medical equipment site. 620 pages. The rework covered 31 pillar pages. Perplexity citations went from 2 to 17 in five weeks. Classic organic traffic gained 14% on restructured pages.

Your action plan: win the decision layer

You don’t need to wait. On e-commerces, I do in this order.

1. Take 20 high-value pages. Not 100.
2. Open each in ChatGPT, Claude, Gemini, Perplexity. Type the target query.
3. Note if the page is cited, and at what position.
4. Rewrite the first block: direct answer in one sentence.
5. Cut the rest into tagged blocks with sub-questions.
6. Add a visible update date.
7. Weave links to sister pages.
8. Measure again 3 weeks later.

The diagnosis takes 45 minutes per page. You do it in-house, or you ask me for a live audit.

The most underestimated point: semantic linking. Without it, AIs see isolated pages. With it, they see a coherent system. The more pages respond to each other, the more AIs trust the site.

Another point: signature and date. Perplexity requires them. Claude and ChatGPT appreciate them. Never publish a page without author and date.

Structured data Product, FAQPage, Article accelerate extraction. For Gemini, it’s a prerequisite. For others, it helps sorting.

Stop continuous flow content. A 1,500-word page with 12 immediate blocks beats five 800-word pages without structure.

Forget « is my content good? ». The only question that matters: does your response block go into extraction in 2 seconds?

AIs don’t care about volume. What they want is clarity. And that builds block by block.

Neil Patel says it too: the decision layer is mechanics, not mystery. The page that gives the answer first wins. The one that tells a story first loses.

Which page of your catalog deserves to be extracted first?

Stéphane Jambu

SEO & AI Engineer

I build growth systems / AI / Neuroscience | 650+ clients · 80 LinkedIn testimonials · 30 years of expertise · 15 years of systems running without me.

Follow on LinkedIn