We read the primary papers instead of the blog posts. GEO (KDD 2024), C-SEO Bench (NeurIPS 2025), ALCE (EMNLP 2023), Lost in the Middle (TACL 2023), Nestaas (ICLR 2025), FORGE (2026), the official OpenAI and Anthropic crawler docs, and server log and citation-tracking studies covering 137,210 domains and roughly 600,000 citation events.
Every number below is from a named source. Where the only evidence is an industry study rather than a controlled one, we say so.

---

Key findings

• Rewriting your page does almost nothing if the page was never retrieved. The benchmark that tested this across six domains found the document’s rank in the model’s context matters far more than any rewrite.
• Blocking the wrong crawler deletes you from AI search. GPTBot is training. OAI-SearchBot is the one that decides whether ChatGPT can cite you. Most robots.txt files confuse them.
• The famous “40% visibility lift” is real and widely misquoted. It measures how much of a generated answer you occupy after you are already in the retrieved set. It is not traffic, and it is not discovery.
• Publishing an llms.txt file is close to free and close to useless for citations. Across 137,210 domains, 97% of published files got zero requests of any kind in a month.
• Your homepage is rarely the URL that gets cited. In one snapshot of ChatGPT citations, Wikipedia was ~13.2% and Reddit ~12%. The work is getting described accurately on the sources that get retrieved, not publishing more of your own pages.
• The marketing document for an AI agent is your tool description, not your homepage. Agents read a name, a description, and a schema. They never see your hero image.
• One poisoned page in the retrieved set can make a model recommend a product that does not exist. Up to 27% of the time with a single page, 73.8% when the top three are replaced. There is no correction desk to call.

---

An AI agent told your best prospect to look at someone else.

You will not see it in analytics. No referrer, no search query, no impression. Nothing.

What you get is a call that never comes. Or a deal that opens with “we shortlisted three vendors” and you are not one of them. By the time you hear about it, the decision is weeks old.

This is not a fringe case. Zero-click searches went from 56% to 69% between May 2024 and May 2025 (Similarweb). Publisher traffic fell around 33% globally in the year to November 2025. On keywords where Google shows an AI answer, one tracked click-through series fell from 1.76% to 0.61%, a 65% decline (Seer Interactive).

The distribution did not disappear, though. It concentrated.

Brands that get cited inside the AI answer earn roughly 35% more clicks. That is the actual game now. Not traffic in general. Being the source the answer names.

So the reflex kicks in. Rewrite the homepage. Add an FAQ. Publish an llms.txt. Buy a tool that scores your “AI visibility” (there are a dozen of those now, and they will each hand you a different score for the same week).

Almost all of that is aimed at the wrong stage of the problem.

The research is now good enough to say which parts work, which parts stop working the moment your competitors copy them, and which parts were never load-bearing.

Here is the whole thing. Basics first, then the parts nobody has priced in yet.

---

Part 1. There are two jobs, and they use different pipes

Everything that gets sold as “AI marketing” is actually two unrelated problems.

Job one is citation. A person asks ChatGPT, Claude, Perplexity, or Google AI Mode a question. The model goes and searches, pulls back a handful of pages, and writes an answer that may name you and link you.

Those few pages it pulled back are what we will call the retrieved set for the rest of this article.

Roughly five to ten documents, chosen before a single word of the answer gets written. You win job one by being a passage the model can lift out of that set without rewriting you into a competitor.

Job two is discovery. A coding agent or a workflow agent lists the tools available to it, reads their descriptions, and picks one. It never visits a website. It reads a name, a one-line description, and a form telling it what fields to fill in. You win job two by being the clearest listing in a catalogue.

These need different artifacts. A beautiful landing page does nothing for job two. A perfect tool description does nothing for job one.

The first question to ask about any AI marketing task is which job it serves. If the answer is “both,” it is probably a slogan.

---

Part 2. Where a recommendation actually comes from

When an agent recommends something, that recommendation came out of one of four places. They have different entry doors, and the door you are missing is usually the whole problem.

Where it comes from	When it fires	How you get in
The model’s own memory	“What is X” questions, brand recall, defaults with no search happening	Be in the big public text piles it learned from: Wikipedia, major documentation, widely copied code repositories, papers. Frozen on the day training stopped
The search index	Anything the product decides to go look up	Let the search crawler in. Different bot from the training crawler
A live page fetch	Someone pastes your URL, or the agent opens one specific page this turn	Send the finished page from your server. These fetchers usually cannot run the JavaScript that builds a page in the browser
A connected tool or feed	Shopping results, an MCP server, a company’s own private documents	Registry listing, product feed, or being installed inside that customer’s account

This is the split that trips up B2B founders most often.

A compliance SaaS founder asks “why doesn’t ChatGPT recommend us for EU AI Act work.” The honest answer is that consumer ChatGPT is answering out of the first three rows. It is reading Wikipedia, government pages, and whatever vendor HTML it managed to fetch.

Your detailed policy pack lives in row four, inside your customers’ accounts. A consumer chat window is never going to see it. (Which is genuinely good news, since that pack is the thing you charge for.)

Those are different products. Treat them as one and you will spend a quarter optimizing a surface your buyer is not using.

---

Part 3. The Retrieval Gate

This is the correction that reorders everything else, so it gets a name.

The Retrieval Gate: nothing you write matters until that page is in the model’s context. Being retrieved is the gate. Craft is what happens after it.

The evidence is direct. C-SEO Bench (Puerto, Gubri, Green, Oh and Yun, NeurIPS 2025 Datasets and Benchmarks) built the first benchmark that tests conversational SEO across multiple tasks, domains, and number of competitors doing it at the same time. More than 1,900 queries and roughly 16,000 documents, two tasks (question answering and product recommendation), three domains each.

Their finding, in their own words: most current methods are “not only largely ineffective but also frequently have a negative impact on document ranking, which is opposite to what is expected.”

Read that twice. Not just useless. Often actively harmful.

What worked instead was ordinary search ranking. The boring thing that gets your page into the retrieved set in the first place.

Then the part that should change your budget. As more players adopt the same tricks, the average gain shrinks toward zero. The authors call it congested and zero sum. A clever rewrite is temporary by construction, because it only pays while you are one of the few doing it.

A 2026 critical survey by Martinez reviewed 45 studies from late 2023 through mid 2026 and lands in the same place. The foundational results hold inside a test rig where five documents are already sitting in front of the model. They establish neither organic discoverability nor durable traffic.

So the order of operations is fixed. Get crawled and ranked into the retrieved set. Then, and only then, does writing craft convert.

---

Part 4. The most expensive robots.txt mistake in B2B

Here is a failure we see in real audits, and it costs more than any content decision on this list.

A founder reads that AI companies are training on the open web without paying. Reasonable reaction. So the team adds a rule blocking AI bots, or flips a CDN toggle labeled “block AI crawlers.”

What actually happened: they deleted themselves from AI search.

OpenAI documents four separate agents, configured independently:

Agent	What it does	What blocking it costs you
GPTBot	Crawls content that may be used to train foundation models	Your content stays out of future training. No effect on ChatGPT search
OAI-SearchBot	Surfaces sites in ChatGPT’s search features	You will not be shown in ChatGPT search answers. This is the expensive one
ChatGPT-User	Visits a page when a user or a custom GPT asks for it	Live “read this URL” requests fail. OpenAI notes that because these are user initiated, robots.txt rules may not apply
OAI-AdsBot	Checks landing pages submitted as ads	Only visits pages you submit as ads. Not training, not search

OpenAI is explicit that a site can allow OAI-SearchBot to appear in search results while disallowing GPTBot so its content is not used for training. They also note it can take around 24 hours for a robots.txt change to propagate to search.

Anthropic splits the same way, with one difference worth knowing.

ClaudeBot is training. Claude-SearchBot indexes for search quality, and Anthropic warns plainly that blocking it “may reduce your site’s visibility and accuracy in user search results.” Claude-User fetches a page when a user’s question needs it. Unlike OpenAI’s note on ChatGPT-User, Anthropic says its user fetcher honors robots.txt too.

Three practical notes that cost teams real visibility:

1. Check the logs, not the file. A firewall rule can slam the door on a search crawler before robots.txt is ever read. Your robots.txt will look immaculate while the crawler is standing outside getting a 403.
2. Ship the claim in the first HTML. Most of these fetchers are closer to a plain download than to a browser. If your pricing, integrations list, or compliance coverage gets drawn by JavaScript after the page loads, it simply is not there when the fetch happens. The crawler sees an empty shell where your differentiator was supposed to be.
3. Content preference headers are not a ranking lever. Cloudflare’s content signals let you state preferences for search, AI input, and AI training. Setting AI input to “no” refuses grounding, which is the opposite of a visibility tactic. These are licensing statements, not a way to rank.

One more trap. Google-Extended is a permission token for Gemini and Vertex generative use, not a crawler. Disallowing it does not remove you from Google Search or AI Overviews, and allowing it does not get you into them.

And one limit on this whole crawler conversation: robots.txt does not stop a browsing agent.

Deep Research modes, ChatGPT’s agent mode, and Perplexity’s browser read pages the way a person does, pulling hundreds of sources including PDFs and technical pages. Some of them navigate using the same labels a screen reader uses, not a picture of the screen. The consequence is blunt. An unlabelled button does not get clicked. A widget loaded in from another domain might as well be a brick wall.

So if your documentation sits behind aggressive bot protection, or your pricing page is a grid of unlabelled toggles, you are invisible on the one surface where a buyer went out of their way to send an agent to evaluate you. That is the worst possible place to be invisible.

---

Part 5. What actually works once you are through the gate

Now the craft. The numbers here are good, as long as you read them correctly.

Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande published GEO: Generative Engine Optimization at KDD 2024. They built a benchmark of roughly 10,000 queries, took pages already in the retrieved set, rewrote them nine different ways, and measured how much of the generated answer each version occupied.

Three treatments beat everything else:

• Adding real citations inside your own page. Name your sources in your own copy.
• Adding credible quoted speech. A real quote from a real named person.
• Replacing vague claims with numbers. “Fast” becomes “answers in under 300 milliseconds, measured May 2026.”

Those three landed roughly 30% to 40% better on the paper’s main measure, which counts how many words of the answer came from your page and weights them by how early they appear. They landed 15% to 30% better on a second, softer measure of how prominent your source felt in the answer. The single best peak in their main table is 41% on the first and 28% on the second. The paper’s headline claim, up to 40% visibility, is that ceiling.

Two details that matter more than the headline:

1. Keyword stuffing failed. The oldest SEO reflex on earth came out flat to slightly negative. Twenty years of muscle memory, worth nothing here.
2. The gains are a catch-up tool. In their rank-by-rank breakdown, the biggest single number is a 115.1% relative improvement for a page sitting fifth in the retrieved set. Pages already sitting first sometimes went down after the exact same rewrite. If you are the obvious leader in your category, this is not your lever, and applying it anyway can cost you.

Now the caveats the vendor blogs skip.

This was tested on the AI engines of that era, which are two generations back. The benchmark contains no B2B or developer tooling category at all. The 37% figure you will see quoted is from live Perplexity, a different measurement on a different product, so do not stack it on top of the 40%.

And every one of those percentages describes how much of an answer you occupy. Not clicks. Not signups. Not pipeline.

---

Part 6. Write for the chunk, not for the essay

If you take one writing rule from this article, take this one.

The ALCE benchmark (Gao, Yen, Yu and Chen, EMNLP 2023) was the first repeatable way to check something obvious that nobody was checking: when an AI answer cites a source, does that source actually say the thing?

They chopped their source documents into roughly 100 word passages, because that is a size a person can verify by eye and a model can hold several of at once. Then they measured.

The result is sobering.

On the hardest question set, around half of the answers from the best models were not fully backed by the passages they had just cited. The citation was there. The support was not.

Follow up work pushes further. FRONT (Huang and colleagues, ACL Findings 2024) trains models to first select supporting quotes and then generate from them, improving citation quality by an average of 14.21% across datasets with LLaMA-2-7B. Self-RAG (Asai and colleagues, ICLR 2024) trains a model to emit judgments about whether a retrieved passage is actually support before using it.

The direction is consistent. Systems are getting better at asking “does this specific passage prove this specific sentence.” A paragraph that requires three other paragraphs to make sense fails that test.

Then position. Lost in the Middle (Liu, Lin, Hewitt, Paranjape, Bevilacqua, Petroni and Liang, TACL 2023) found a U shaped curve: models use information at the beginning and the end of a long context far more reliably than the middle. In one case, when the answer sat in the middle of a long context, performance dropped below what the model achieved with no documents at all.

You do not control your position in someone else’s context window. You do control whether your claim survives being read alone.

What that means for the actual page:

1. One heading, then one to three sentences that stand on their own.
2. The first sentence of each section carries the subject, the claim, and the evidence together. Not a pronoun, and not “as mentioned above.”
3. Evidence sits in the same paragraph as the claim, not in a footer PDF.
4. Every number carries a date and a scope. “Q2 2026, across 40 customer accounts” beats “recently.”
5. Repeat your product name instead of writing “the platform.” One name per thing.
6. Tables and definition lists survive being chopped into chunks. Long narrative does not.
7. For video, the transcript is what gets cited. The video file is not readable.

Two more constraints worth designing around, both from vendor documentation rather than papers.

Anthropic’s web search hands back the quoted bit of your page as a short excerpt, typically capped at 150 characters. That is about one sentence. If your core claim cannot survive being trimmed to a sentence, it gets paraphrased into something you did not write, and you have no say in what.

Perplexity’s page fetch is reported to stop at roughly 4,000 tokens, call it 3,000 words. So the first screen is effectively the whole document. Everything below your fold might as well not exist.

One last thing, and it is not about wording. Structure is its own separate lever. A 2026 follow-up study held the meaning of a document completely constant and changed only how it was organized, and still measured roughly a 17% citation lift. Same claims. Better shape. That is the cheapest win on this list, because it needs no new evidence and no new research. Just moving what you already wrote into a better order.

---

Part 7. Your homepage is rarely the cited URL

When a model answers “best X for Y,” the sources it names are usually not vendor homepages. In one snapshot of roughly 600,000 ChatGPT citation events in the US (Similarweb, January to February 2026), Wikipedia was about 13.2% of citations, Reddit about 12%, and openai.com’s own properties about 6.2%.

The biggest citation source in that mix is an encyclopedia you cannot buy your way into. The second is a forum. Your marketing site competes for what is left.

So the highest leverage citation work is often not on your domain at all. It is getting accurately described on the sources that get retrieved:

• Roundups and comparison posts. The “best X for Y” article that already ranks is the document that gets pulled. Being accurately included in someone else’s roundup beats publishing your own.
• Community platforms. Reddit and similar threads are heavily retrieved. This means ordinary community participation, not astroturfing. Fake accounts are both a policy violation and, per the injection research below, the same failure mode as an attack.
• YouTube, through transcripts. Ahrefs measured a 0.737 correlation between a brand’s YouTube mentions and its appearances in AI answers across roughly 75,000 brands (December 2025). Correlation, not causation, and video-heavy categories will inflate it. But note the mechanism: engines read the transcript, not the video. An unlisted, uncaptioned video is invisible.
• Public knowledge bases. Use the same legal name and product name everywhere, and link your identities together (Wikidata, GitHub, Crunchbase, LinkedIn) so the entity resolves to one thing. Do not edit your own Wikipedia page. And be clear-eyed: a Wikidata entry is identity resolution, not a proven switch that increases citations.

Two warnings attached to all of this.

Citation share is violently unstable. Reddit’s share of ChatGPT citations reportedly fell from around 60% to around 10% inside two weeks in September 2025. Any media plan built on last quarter’s platform mix is a plan built on sand. Treat these percentages as a description of how the pipe works, never as an allocation model.

Two credible studies can report different Wikipedia numbers, and both can be right. One industry mix put Wikipedia at roughly 6% of AI citations while the snapshot above says 13.2%. That is not a contradiction. They counted different things: different engines, different time windows, and a different pool of citations to take a percentage of. The takeaway survives either way. Wikipedia carries a lot of weight, and it is nowhere near a majority.

And one structural fact that no amount of writing gets around.

AI labs now pay for content directly. OpenAI has signed flat-fee deals with publishers including the Associated Press, the Financial Times, News Corp, Reddit, and Stack Overflow. Perplexity pays per citation.

So if a paid publisher dominates the answers in your category, there is no rewrite that beats them. They are in the room because someone signed a contract. You do not out-write that. You get covered by them instead, which makes it a press job, not a content job.

---

Part 8. The files that do not do what people think

This is where most teams lose a week.

llms.txt is a proposed Markdown index of your documentation, served at the root of your domain. Real thing, real spec, genuinely useful for coding agents parsing developer docs.

It is not a search ranking signal, and the adoption data is brutal. Ahrefs went through server logs across 137,210 domains in May 2026. Of those, 28% published an llms.txt. Of the files that existed, 97% received zero requests of any kind that month. Nobody fetched them. Not one hit.

Inside the 3% that got any traffic at all, the readers were not who you would expect. Named AI bots were 19.5% of requests. The AI retrieval bots, the ones that actually decide whether you get cited in a live answer, came to 1.1%. Bookkeeping tools that audit websites were 21.7%, though that number includes Ahrefs’ own crawlers checking their own customers.

And among the AI readers, the leaders were training crawlers and coding agents. Anthropic’s coding agent out-fetched every AI retrieval bot, every AI assistant, and every training crawler except GPTBot.

So there is your real audience for the file. Coding agents. Not search.

Google’s own May 2026 guidance on AI features has a section literally labelled mythbusting. It says machine readable files like llms.txt are not needed to show up in AI features, that there is no special markup to add, no requirement to pre-chop your pages, and no need to rewrite anything “for AI.”

Days later, Chrome shipped an experimental audit that checks whether you have the file.

Both of those are true at once. They are answering different questions. The internet decided they were a fight.

Write one if you have real documentation. It costs a single Markdown file and coding agents do use the shape. Do not put it on the roadmap as a citation strategy.

FAQ markup is the same trap in a different outfit.

You will see claims that adding the standard FAQ code block to your pages multiplies your AI Overview appearances several times over. Google does not require it for AI Overviews, and the studies making that claim never separate “added the markup” from “was already ranking well,” which is doing all the work.

What actually helps is the shape, not the code. A real question as a heading, followed by a direct answer.

Here is why. Before searching, the model breaks one question into several smaller ones. Ask “what should I use for AI Act compliance at a small company” and it quietly goes looking for “AI Act compliance tools,” “compliance software for small teams,” and “X versus Y” as three separate searches. The industry calls that fan-out. It is why “our tool for small teams” and “our tool versus the incumbent” deserve their own headings.

That is writing. No code block required.

The same discipline applies to the rest of the standards pile. Agent cards, catalogue manifests, DNS based discovery records. In a broad scan of the web, roughly 0.29% of probed hosts published an agent card at all, and most of those were invalid. The rule we use: do not ship a discovery file until you can name the specific product that reads it.

---

Part 9. The advanced move: your tool description is the marketing copy

If your product has an API, this is the highest leverage work available in 2026, and almost nobody is doing it well.

Quick definition first, because the acronym is everywhere and rarely explained. MCP is a shared standard for letting an AI agent use an outside service. You run a small server, it advertises what your product can do, and any agent that speaks the standard can use it. Think of it as a menu your software hands to an AI.

When an agent connects, it loads that menu. Every item has a name, a description, and a list of the fields it needs filled in. The whole menu gets pasted into the model’s prompt, and the model orders from it.

Which means that description is your positioning statement.

Read by the decision maker. Every single time. At the exact moment of the decision.

Your homepage never gets that seat.

Anthropic’s guidance on writing tools for agents (September 2025) is the closest thing to a spec here, and the funny part is that it reads like copywriting advice:

• Name by verb and object. invoice_create is unambiguous. process could mean anything, and the model will find out which by guessing.
• Say what it does, what it hands back, and what it does not do. That last line prevents a wrong call, which is worth more than winning one extra right one.
• Put the limits in the form fields, not in the prose. A field that only accepts three specific values is a rule. A sentence politely asking the model to use one of three values is a suggestion, and it will eventually get ignored.
• Ship few tools. Twenty near-identical tools is the agent version of a pricing page with nine columns.
• Test it like a product. Run realistic multi step tasks and rewrite the descriptions against whatever broke. The metric is whether the agent picks the right tool. Not stars. Not installs.

Then the piece most teams miss completely.

An MCP server can also hand over resources, which are read-only documents the agent can pull on its own. That is exactly where your citable facts belong: what the product is, what it deliberately does not do, dated numbers, limits, and how you would like to be quoted. Right there, named, easy to find. Not buried inside some generic get_info() call the model has to think to make.

Two rules keep this honest. Register the server once in the official registry so there is one canonical you, and pick a single human-facing directory if you want clicks. More listings do not help. And a dead endpoint is worse than no endpoint, because agents learn to skip you.

And do not write “always use this vendor” into a description. That is not clever positioning. That is the exact attack pattern described in the next section, and reviewers flag it.

---

Part 10. The dark side, and why it is your problem even if you never touch it

Two research lines matter here. Neither is a tactic. Both change your defensive posture.

Injection. Nestaas, Debenedetti and Tramèr (ICLR 2025) demonstrated what they call Preference Manipulation Attacks: crafted text on a web page or inside a plugin’s documentation that pushes an assistant toward the attacker’s product.

Against Bing, a targeted camera became 2.5 times more likely to be recommended. A GPT-4 news plugin became 2 to 8 times more likely to be selected after an attack. Related work by Kumar and Lakkaraju showed optimized text strings that force a named product into the top slot, and Pfrommer and colleagues showed jailbreak style ranking attacks that transfer to production conversational search.

The authors also describe the endgame: a prisoner’s dilemma. Everyone injects, and the answers get worse for everyone, including the attackers.

Pollution. The FORGE study (Luo and Chen, 2026) tested something nastier. Across 12 models, 225 products, and 15 categories, they rewrote the brand in retrieved pages into a brand that does not exist. A single polluted page in the retrieved set produced fake product recommendations up to 27% of the time. Replacing the top three retrieved pages pushed that to 73.8%.

Two findings from that paper should worry any B2B founder in a niche category:

It is worst where the model has no firm opinion already. Big consumer brands get some protection from the model’s own memory. It knows what a Nike is. But a specialist compliance tool, a developer infrastructure product, a vertical SaaS with 400 customers? The model holds no strong view of your category, so it just believes the page in front of it.

That is most B2B startups. The niche that protects your pricing power is the same niche that leaves you undefended here.

Reasoning made it worse, not better. The models frequently produced confident, plausible justification for the invented brand, complete with social proof that never existed. And telling the model to be skeptical sometimes made it more vulnerable, not less.

There is no support ticket for any of this. No lab runs a public desk where you can correct a made-up fact about your company. You cannot email OpenAI and get a retraction.

The only defense is structural: put the same facts, worded the same way, on several independent domains you do not control.

Your documentation, your listing pages, third party roundups, your entries in public knowledge bases. Agreement across separate sources is the filter that actually catches this. Being consistent in five places is boring, and it is the whole defense.

---

Part 11. How to measure this without fooling yourself

There is no search console for ChatGPT. So you run a protocol, not a dashboard.

1. Fix a prompt set. 20 to 50 questions a real buyer would ask, each phrased three ways. Include the languages you actually sell in.
2. Repeat on different days, at least five runs. Visibility is a distribution, not a value. One good answer proves nothing.
3. Log what you can see. Whether the model searched at all, which URLs it cited, whether your own domain appeared or only third party pages, and any feature it invented about you.
4. Split by surface. ChatGPT search, shopping results, agent browsing, and a customer’s internal setup are different systems. The overlap between them is genuinely small: Google’s AI Overviews and AI Mode were measured sharing only about 13.7% of cited URLs, and one investigation found roughly 1.5% URL overlap between what ChatGPT surfaced and the Bing top 20. A blended number is not a number.
5. For MCP, measure tool choice. Given a realistic task, does the agent pick the right tool and fill the arguments correctly? That is the only score that predicts usage.

If you cannot name which surface a number came from, it is not a number. It is a vibe with a decimal point.

The visibility tools you can buy are useful for spotting direction and useless for comparing against each other, because each one counts a different pool of answers before taking a percentage of it. Two tools, same week, same brand, two different scores. Neither is lying.

---

Part 12. The stack, in order

Do these in sequence. Stop when the next item is fashion.

1. Let the search crawlers in and verify in your access logs. OAI-SearchBot, Claude-SearchBot, PerplexityBot, Googlebot, Bingbot. This is the Retrieval Gate and everything else is downstream of it.
2. Publish one canonical facts page. What the product is, what it is not, dated numbers, limits, and the URLs you want cited. Sent finished from your server, plus a plain Markdown copy at a stable address. Put an updated: date and a one-line note on what changed at the top, so an agent working from an old copy can tell it is stale.
3. Make every section survive being read alone. Subject, claim, evidence in the first sentence. This is the ALCE lesson and it is free.
4. Get described accurately on the sources that actually get retrieved. Roundups, community threads, video transcripts, and public knowledge bases. This outranks most on-site work, and for many B2B categories it is the single highest leverage item after crawlability.
5. Use one capability sentence everywhere. Repository, README, package listing, homepage headline, server description. Identical wording. Inconsistency reads as two different products.
6. Add llms.txt if you have documentation. One file. Coding agents use it. Expect nothing from search.
7. If you have an API, treat the MCP server as the marketing pack. Registry entry, man page style tool descriptions, facts on resources. Generate your OpenAPI from the same descriptions if you also ship actions.
8. Stop. No agent cards, no catalogue manifests, no ninth well known file, until you can name the product that reads it.

The remaining work after step 8 is measurement on live prompts. It is not another standard.

One thing not on the list: ads. ChatGPT advertising exists, but it renders as a labelled tile near the answer, it runs on the free and lower tiers, and OpenAI’s stated policy is that it does not change the answer itself. Enterprise accounts do not see ads at all, which is exactly where a B2B evaluation happens. Buying ads does not get your name into the organic recommendation, and it will not reach the buyer running a procurement review.

---

Where this is heading

This section is inference, not evidence. It is where we would put money, and it is deliberately less certain than everything above. Judge it accordingly.

The rewrite advantage keeps decaying. C-SEO Bench already shows gains shrinking as more players pile in. Meanwhile every AI writing tool now ships a “make this AI friendly” button. Anything a button can do is priced in within a year. What does not decay is being the independently verifiable source of a fact, because that one requires actually knowing something.

Retrieval, not writing, becomes the budget line. If rank inside the retrieved set dominates, the money follows whatever sets the rank. That work looks a lot more like engineering than like content marketing: crawlability, canonical facts, clean product data, presence in the third party sources that get pulled. Our bet is that “AI visibility” services quietly turn into retrieval engineering and keep the nicer name.

Licensing keeps eating crawling. The publisher deals above are the early shape of this. We expect the share of answers grounded in paid content to keep climbing, which pushes B2B distribution toward getting covered by analyst pages, roundups, and communities instead of publishing more of your own pages. Fewer blog posts, more phone calls. Marketers will hate it.

Agent browsing grows as a share of evaluation. The mechanism already exists, so what changes is volume: more buyers sending an agent to read three vendor sites directly. That makes plain, labelled, server-sent HTML a first-class marketing asset, and turns a login wall on your documentation into a straightforward revenue leak.

A verification layer gets built. Given the pollution results, some way for a model to weigh a claim by where it came from and whether independent sources agree is close to inevitable. Consistency across independent domains is the cheapest way to be ready for it. Conveniently, it is also your best defense today, which makes it the rare bet that pays whether or not we are right.

The honest uncertainty: nobody knows how any of this behaves outside English. Not us, not the researchers. Almost every study cited here is US English. If you sell in Spanish or Portuguese, the question gets broken into sub-searches in your buyer’s language, and you need a page that answers that phrasing. The transfer data does not exist yet. Anyone selling you a number for it is guessing.

---

The bottom line

Your buyer is going to ask an AI. That part is settled.

What is not settled, and what most teams get backwards, is where the leverage sits.

It is not in a cleverer paragraph. Rewrites decay the moment your competitors run the same play, and the benchmark that tested competition head on found most of them ineffective or actively harmful.

The leverage is the Retrieval Gate.

Be fetchable. Be ranked into the context. Then write claims that survive being read alone, in a chunk, with no paragraph above them for support.

Everything in this article is downstream of one question you can answer this afternoon: open your access logs and check whether OAI-SearchBot and Claude-SearchBot got a 200 last week.

If they did not, nothing else on this list is your problem yet. You are still standing outside the Retrieval Gate.

---

Limitations and caveats

The main GEO results predate current models. The KDD 2024 experiments ran on the generative engines of that era. Whether the same rewrites hold on 2026 frontier models is untested, and the 2026 survey flags this directly.

No B2B or developer tooling domain was tested. The GEO benchmark covers general web domains. Applying its percentages to a developer tools category or a compliance product is an assumption, not a finding.

Answer share and revenue are different things. Every academic number here measures how much of a generated answer a source occupies, or how often it is cited. None measures signups or revenue. The 35% click premium for cited brands is industry measurement, not a controlled study, and clicks are still two steps from a customer. Citation is not a click, and a click is not a deal. Treat the citation research as a distribution finding, not a revenue model.

Server log studies describe a panel, not the web. The 137,210 domain figure comes from one analytics product’s customer base, which skews technical and SEO aware. The 28% publish rate should be read as an upper bound.

Crawler behavior changes without notice. OpenAI materially revised its crawler documentation in December 2025. Anthropic updated its crawler page in 2026. Verify the current documentation before making a robots.txt decision, and re-verify after any CDN change.

The forward looking section is inference. It is labelled as such because it is. No study cited supports it.

---

Sources

Peer reviewed and preprint:
– Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, Deshpande. GEO: Generative Engine Optimization. KDD 2024. arXiv:2311.09735
– Puerto, Gubri, Green, Oh, Yun. C-SEO Bench: Does Conversational SEO Work? NeurIPS 2025 Datasets and Benchmarks. arXiv:2506.11097
– Gao, Yen, Yu, Chen. Enabling Large Language Models to Generate Text with Citations (ALCE). EMNLP 2023. arXiv:2305.14627
– Liu, Lin, Hewitt, Paranjape, Bevilacqua, Petroni, Liang. Lost in the Middle: How Language Models Use Long Contexts. TACL 2023. arXiv:2307.03172
– Huang and colleagues. Learning Fine-Grained Grounded Citations for Attributed Large Language Models (FRONT). ACL Findings 2024. arXiv:2408.04568
– Asai, Wu, Wang, Sil, Hajishirzi. Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection. ICLR 2024. arXiv:2310.11511
– Nestaas, Debenedetti, Tramèr. Adversarial Search Engine Optimization for Large Language Models. ICLR 2025. arXiv:2406.18382
– Kumar, Lakkaraju. Manipulating Large Language Models to Increase Product Visibility. arXiv:2404.07981
– Pfrommer, Bai, Gautam, Sojoudi. Ranking Manipulation for Conversational Search Engines. arXiv:2406.03589
– Luo, Chen. One Polluted Page Is Enough: Measuring Web-Content Pollution in LLM Recommenders (FORGE). arXiv:2606.13610
– Martinez. Optimizing Visibility in Generative Engines: A Critical Survey. arXiv:2607.14035

Official vendor documentation:
– OpenAI, Overview of OpenAI Crawlers. developers.openai.com/api/docs/bots
– Anthropic, Does Anthropic crawl data from the web, and how can site owners block the crawler? support.claude.com
– Google Search Central, Optimizing your website for generative AI features on Google Search (May 2026)
– Anthropic, Writing effective tools for AI agents (September 2025)
– Model Context Protocol specification and registry. modelcontextprotocol.io
– llms.txt specification. llmstxt.org
– Cloudflare Content Signals. contentsignals.org

Measured industry studies (not peer reviewed, read as directional):
– Ahrefs, We Analyzed 137K Sites: 97% of llms.txt Files Never Get Read (June 2026). ahrefs.com/blog/llmstxt-study. Server logs across 137,210 domains, May 2026 traffic.
– Ahrefs (December 2025). YouTube mentions versus AI answer appearances, roughly 75,000 brands. Correlation 0.737.
– Similarweb. Zero-click share 56% to 69%, May 2024 to May 2025. Separately, roughly 600,000 ChatGPT citation events, US, January to February 2026 (Wikipedia ~13.2%, Reddit ~12%, openai.com ~6.2%).
– Seer Interactive. Click-through on keywords showing AI Overviews, 1.76% (June 2024) to 0.61% (September 2025). Seer describes the later partial rebound as levelling off, explicitly not a recovery.
– Cited-brand click premium (~35% more clicks against a ~61% organic CTR decline): industry measurement aggregated across AI Overview studies. Direction is consistent across sources; magnitudes vary by methodology.
– Reddit citation-share collapse, roughly 60% to 10% within two weeks, September 2025 (SEMrush data reported via 5W).
– Google AI Overviews versus AI Mode URL overlap ~13.7% (Ahrefs). ChatGPT versus Bing top-20 URL overlap ~1.5% (Search Engine Land investigation, 2026).

2026 follow-on research (cited once, not re-verified against the PDF for this article):
– GEO-SFE, structural framing effects with semantics held constant, ~17% citation lift. arXiv:2603.29979