Jordan Neville, founder.

The rest of this page is what I think about generative and answer engine optimization, and why almost everything being sold under that name right now is worth very little.

Start with the test

Open ChatGPT and ask it to recommend a company in your category. Count the names. Usually two, sometimes three. That is the whole shortlist. No page two, no ad slot, and nothing in your analytics will tell you that you were left out. No impression to miss. No click to lose. Just absence.

Run it again tomorrow and you will get different names to the same question.

That is where the real work starts, and where most of the industry stops.

Why this is not a side channel anymore

Forrester's 2026 Buyers' Journey Survey of roughly 18,000 global business buyers found that 94 percent used AI during their most recent purchase, 55 percent compared vendors inside AI tools, and 47 percent built internal business cases before contacting a single vendor. G2's April 2026 research found that 69 percent of B2B software buyers chose a different vendor than they originally planned based on what a chatbot told them, and a third bought from a vendor they had never heard of before.

On the consumer side it is further along than most brands realize. Adobe's March 2026 data showed AI-referred traffic converting 42 percent better than non-AI traffic, reversing a deficit of 38 percent recorded a year earlier. That traffic is smaller and worth more, because the model already did the qualifying.

Then there is the layer nobody has priced in yet. Agent-mediated purchase. Three protocols are being negotiated in the open right now: ACP from OpenAI and Stripe for checkout execution, Universal Commerce Protocol from Google and Shopify for the wider commerce journey, and AP2 for payment authorization, which Google handed to the FIDO Alliance in April 2026. The consumer-facing checkout button inside ChatGPT was built, shipped, and retired inside about five months, and the protocol underneath it survived. That is the tell. The interface layer is volatile. The machine-readability layer underneath it is where the durable advantage sits.

And access is not neutral. Amazon blocks OpenAI's crawlers in robots.txt, which means Amazon listings cannot surface in ChatGPT shopping results at all. One line in a text file decided a category. Most brands have never read their own.

The landscape, as it actually behaves

Eight platforms matter today and they do not share an index, a crawler, a refresh cycle, or a definition of authority.

ChatGPT decides per turn whether to retrieve at all or answer from weights that were frozen months ago, and the answer you get differs by model version, by whether reasoning is on, and by whether the user is on a fast or deep research mode. Google AI Mode and AI Overviews decompose one question into a fan of sub-queries you never see, run them in parallel, and synthesize the result, so the thing being answered is not the thing the buyer typed. Perplexity retrieves live and cites aggressively, which makes it the most forgiving surface and the most misleading one to test on if it is the only one you test. Copilot inherits Bing's index and its own grounding rules. Gemini is entangled with Google's Shopping Graph and account-level personalization. Grok weights a social corpus most brands have no presence in. Claude, Meta AI, and DeepSeek each behave differently again, and enterprise deployments of all of them run inside private tenants with their own retrieval.

Underneath the platform layer, the mechanics decide everything.

• Whether the crawler can reach you, and which one. GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot and Google-Extended are separate permissions, and training access is not the same as retrieval access.
• Whether your page chunks cleanly, because retrieval does not read your page, it reads a fragment of it, and a fragment that only makes sense in the context of the paragraph above it will never be extracted.
• Whether the model resolves you to the right entity or blends you with a company that shares half your name.
• Whether independent sources corroborate what you say about yourself, because self-published claims carry almost no weight and third-party corroboration carries most of it.
• Whether the index that surfaced you yesterday refreshed overnight.

Then the buyer's own state.

• Their stored memory of what they told the model last month.
• Their custom instructions, their prior chats, their location down to the metro, their device, their language, their account age.
• Where they are in the conversation, because turn one and turn four are different games and the recommendation usually lands around turn three.
• The exact phrasing they used, since ‘best,’ ‘top,’ ‘who should I hire,’ and ‘who do people actually like’ pull different sources.
• The follow-up they asked afterward, which is where the real shortlist forms and which almost nobody tests.
• Whatever experiment the platform is running on that account this week.
• In regulated verticals, a conservative filter that suppresses named recommendations entirely.

Multiply all of it. That is your visibility surface. It is not a ranking. It is a probability distribution, and it has to be measured like one.

What is being sold, and why it fails

Three versions of this exist in the market right now.

All three share a deeper problem, which is that the tactics being sold are often the ones the platform owners have publicly said do not work. Google's May 2026 documentation stated that llms.txt files, content chunking, AI-specific rewriting, and special schema are not required for visibility in its generative features. Half the checklists in circulation are selling exactly that list. Knowing which surface a tactic applies to, and which platform's stated behavior contradicts it, is most of the job.

How I see it

You cannot control most of the die. Anyone telling you otherwise is selling something.

You can influence a specific subset, and that subset is smaller and more boring than the industry wants to admit. It reduces to whether the machine can reach you, parse you, identify you correctly, and find credible independent sources that agree with what you say about yourself. Everything past that is theater dressed as strategy.

Measuring it honestly requires controlled conditions, clean sessions with no inherited memory, sample sizes large enough to separate signal from variance, prompt sets built around real buyer stages rather than keywords, a defined competitor set, and provenance behind every figure reported. Every number should be traceable to the run that produced it. That standard is the entire difference between an audit and a screenshot, and it is why so few people do this properly.

Because this is zero sum. Every prompt has winners. If a competitor is named and you are not, you were not overlooked. You were displaced. Share of voice inside answer engines is a real, contested, measurable position, and it compounds, because the corroboration that earns you a mention next year is being published or not published right now.

I have spent the last several years building and running that methodology across every major platform, for brands in regulated verticals, in crowded categories, and in categories the models barely understand yet. Some of what I have found contradicts fifteen years of accepted practice, including things I believed six months ago.