What AI search taught me about SEO, AEO, GEO and CRO

The chart that should worry a B2B website team first is not the traffic report. Traffic can look fine. It is how often AI engines name the brand when buyers ask about its category, sitting right next to how often they name its biggest competitor.

A site can rank on Google and still not be the answer. Those are two different jobs, and most of what I have learned about AI search comes from that gap.

This is the version I would give a peer running an in-house B2B website: what SEO, AEO, GEO and CRO each do, how to measure the AI side without fooling yourself, and the changes that are worth making first. No secret sauce, just the things that held up.

Four letters, one job

It helps to name the jobs before arguing about tactics.

• SEO gets the page found and ranked in classic search.
• AEO, answer engine optimization, shapes the page so an answer box, a voice assistant or an AI Overview can lift a passage that answers the question.
• GEO, generative engine optimization, aims for the citation: when ChatGPT, Gemini, Claude or Perplexity answer a question about your category, do they name you, and do they link to you?
• CRO, conversion rate optimization, makes sure the visit all of that earns turns into something: a demo request, a download, a conversation.

They hand the visitor to each other. A gap anywhere leaks the whole funnel, which is why I stopped treating them as four projects.

Measure it like an experiment, not a vibe check

The fastest way to fool yourself about AI visibility is to type a question into ChatGPT, see your brand, and call it a win. Ask again tomorrow and the answer can change. Models get updated, answers are generated fresh each time, and your own history can colour what you see.

What has worked for me is treating it like an experiment:

• Start from real questions. Pull them from Search Console queries, People Also Ask, forums and Reddit threads, review sites and, most underrated of all, your sales team. The questions buyers ask a salesperson are the ones they now ask a chatbot.
• Freeze the benchmark. Pick the questions that matter and keep them fixed, so this month compares with last month.
• Ask more than once. Sample each answer several times and record the pattern, not one lucky reply.
• Separate mentions from citations. Being named in an answer and being linked as a source are different results, and they move for different reasons.
• Report on a rhythm. I read the tool data weekly and turn it into a short change report and a ranked list of what to fix next. When leadership can watch the line move, the work stops being a side project.

Write for the question, not the org chart

Most company websites are organised the way the company thinks: by product name. Buyers do not search for your product names. They search for the problem.

So I map the questions into a small set of buyer intents and build a content hub for each one. The hub answers the question a buyer actually types, in plain language, then links down into the product detail for the people who want it. The product pages do not go away. They just stop being the front door.

Then go back through the pages you already have. This is the unglamorous part, and it is where a lot of the gains hide. New pages get the attention, but old pages carry the links and the history. Rethinking an existing page is often worth more than publishing a new one.

Make every page easy to quote

Answer engines lift passages, not pages. Two habits make the biggest difference to how quotable a page becomes:

• Key takeaways at the top. A short summary under the title, in plain sentences, so the answer sits in the first screen instead of paragraph nine. You are reading a post that does it.
• Stats that survive being lifted. A specific figure with its source and its context, written as one sentence that still makes sense on its own.

FAQs deserve the same care. Write the questions the way people ask them and make every answer stand on its own. Google stopped showing FAQ rich results for most sites back in 2023, so this is not about stars in the search results any more. It is about giving engines clean question-and-answer pairs to work with.

Borrowed credibility, delivered fast

AI engines lean on third parties. An analyst report, a review site or a respected publication saying something about you counts for more than your own homepage saying it.

So when analyst coverage lands, treat it as urgent. Reports like these tend to wait in a queue while everyone debates the landing page, and every week a report is not live is a week the engines cannot connect it to you.

The same goes for gated content. A PDF behind a form is close to invisible to an answer engine. If something is worth quoting, give it a real web page that can be read, indexed and cited, and keep the form for the deeper material.

Check the plumbing nobody looks at

A surprising amount of AI visibility is decided before anyone reads a word.

• Let the right crawlers in. Check your robots.txt and your CDN. Cloudflare began blocking AI crawlers by default for new sites in 2025, so a site can be shut out of AI search without anyone deciding that on purpose.
• Serve the answer in HTML. Most AI crawlers, including the ones from OpenAI, Anthropic and Perplexity, fetch the page without running its JavaScript. If your answer only appears after a script runs, those crawlers never see it.
• Make the schema agree with the page. Structured data that contradicts the visible text does more harm than having none.
• Keep llms.txt in proportion. It is a plain-text map of a site for language models, cheap to add and harmless. Just do not expect much from it yet: Google says it has no effect on Search, and server logs show AI crawlers rarely ask for it.

Then automate the upkeep. Meta descriptions, image alt text, schema and site audits are exactly the kind of work that quietly rots while everyone is busy with the next campaign. Tools built inside the website can handle them, so the basics stay done.

Conversion closes the loop

Here is the part AI search conversations tend to skip. Being cited is not the goal. Being chosen is.

A visitor who arrives from an AI answer has often done much of their research in the chat already. The page they land on has to pick up where the answer left off, with a clear next step. A few things I always come back to:

• Watch before you change. Heatmaps and session recordings show where people stop, long before a test does.
• Test the calls to action. Placement and persistence matter more than wording more often than you would think. Test one change at a time against the current version.
• Never lose a lead in the plumbing. A form that fails quietly when the CRM is down is the most expensive bug on the site. Hold every submission until the CRM confirms it.
• Get ready for agents. AI assistants are starting to fill in forms for people. I wrote about making forms agent-friendly with WebMCP in the other post on this blog.

It is also worth separating AI referrals in your analytics, so visits from ChatGPT, Perplexity, Gemini and Claude show up as their own source instead of disappearing into referral traffic. You cannot improve what you cannot see.

One loop, one owner

The biggest lesson is not a tactic. Most companies split this work between an SEO agency, a content team, a developer queue and someone who owns conversion, and every hand-off is a place where things leak.

When one person or one tight team owns the whole loop, a fix to a template does not break the schema, a new hub does not ship without a call to action, and the measurement actually feeds the next piece of work. That is less exciting than a new tactic, and it is the reason the tactics work.