AEO is the Wild West, which is exactly why it is worth riding into now

Every few years the search industry discovers a new surface, declares it lawless, and immediately sells a map of it. The map is always wrong in the same way: it describes the tactics that worked last month and calls them rules.

Answer engine optimization is in that phase now. The difference is that this time the surface decides what customers are told about your inventory, your pricing and your service department, in sentences your store never wrote.

The first era never had rules either

Between 1998 and 2003, nobody knew how ranking worked. There was no documentation worth the name, no reliable analytics, and no way to tell whether a tactic worked or the index simply moved. So the industry invented doctrine from experiments: keyword density percentages, meta keyword lists, doorway pages, reciprocal link schemes, and a hundred confident explanations of the Google Dance.

Some of it worked, briefly. Almost none of it aged. The Florida update in 2003 removed a large share of it overnight, and the people who survived were not the ones with the cleverest trick. They were the ones who had built crawlable sites, original content and real links, and had measured what happened.

Read the AEO vendor decks in 2026 and you are looking at the same document with different nouns.

What the Wild West actually looks like

Four claims are currently sold as settled fact. None of them are.

llms.txt is a standard. It is a proposal, published in 2024, for a markdown file that summarises a site for language models. It is a sensible idea and it is worth an afternoon. It is not a standard, and nothing publicly available shows that a major consumer answer engine reads it before deciding what to cite. Publish it because it costs almost nothing, not because it moves a number.

An "AI visibility score" measures something. Dozens of tools now produce a single number for how visible a brand is to models. Each uses its own prompt set, its own model, its own sampling rate and its own definition of a citation. Two tools will give you two different scores for the same domain on the same day, and neither publishes an error bar. A dashboard that cannot reproduce itself is a sales instrument. Ask any vendor for the methodology and the sample size before you ask for the invoice.

Volume fixes it. The industry has already produced the AEO equivalent of keyword stuffing: thousands of thin pages engineered to sound like an answer. Models do not reward padding. They quote sources that state something specific, attributable and current. Ten pages of genuine primary data will out-cite a thousand pages of plausible filler, and the filler carries the same risk it carried in 2004.

Citations can be guaranteed. They cannot. Ranking in a generative answer is selection at inference time across sources the model retrieved, weighted by a model you do not control and cannot inspect. Any agency promising placement is selling you a screenshot.

The invariants: what already behaves like a rule

Strip the theatre and there are maybe six things that hold across engines and across model updates. They are unglamorous and they are the whole job.

1. Explicit crawler access. Decide, per bot, whether you want to be retrieved. GPTBot, ClaudeBot, PerplexityBot and Google-Extended are all named in robots.txt conventions, and a site that blocks them has removed itself from the answer layer to protect content it never monetized anyway. Audit the file quarterly and know what each directive is doing. Blanket-blocking AI crawlers is a strategic choice, not a default.

2. Server-rendered content. Several AI fetchers retrieve raw HTML without running JavaScript. A dealer site whose service pricing, hours or inventory copy only exists after hydration is, to those fetchers, an empty page. Open your most important URL with scripts disabled and read what is actually there.

3. Structured data and consistent entities. Schema.org markup is not a ranking trick, it is a disambiguation mechanism. If your organization, locations, service offerings and vehicle data are machine-readable and identical everywhere they appear, a model has something to reconcile. If your name, address and phone differ between the site, the GBP listing and a directory, you have given the model a reason to prefer someone else.

4. Primary data and quotable definitions. The most citable asset on the internet is a specific number nobody else has published, stated plainly with its method and date. In automotive that is a real cost-to-own table, real service interval data, a real days-to-turn series from your own DMS. Models quote sources that say something checkable.

5. Semantics, dates and authorship. Clean heading hierarchy, one topic per page, a visible publication and update date, and a named organization behind the claim. All of it was best practice before anyone said "answer engine," and all of it still decides which page a retrieval step can trust.

6. Presence where retrieval happens. Models do not only read your domain. They read forums, review platforms, comparison sites and community threads. Being described accurately off-site is now part of on-site work, which is a governance problem more than a content one.

Measurement, honestly

There is no answer-engine equivalent of Search Console. Anyone who claims otherwise is describing a scraper.

What you can actually measure:

• Crawler hits in server logs. GPTBot, ClaudeBot, PerplexityBot and friends leave user-agent strings. Count them by section. This tells you what is being retrieved, which is upstream of being cited.
• A fixed prompt panel. Fifty prompts you care about, written down, run monthly by a human in a clean session, recording whether you were cited, quoted, or absent. Boring, repeatable, and more useful than any score.
• Referral and branded search lift. Chat surfaces often strip referrers, and in-app browsing passes nothing at all, so chat referrals are a floor rather than a total. Pair it with branded search volume and direct traffic in the metro.
• Assisted conversion. Sessions that touched an AI referrer and converted later. Small sample now, real direction.

Write the panel down and keep the raw outputs. In two years that archive will be worth more than any tool subscription, because it will be the only evidence of how your visibility actually moved.

A ninety-day sequence that makes sense

1. Crawler audit. What is allowed, what is blocked, what is being hit. One page, signed off.
2. Render check. Your twenty highest-value URLs with JavaScript off. Fix what is empty.
3. Entity pass. Organization, location, service and product schema, plus name, address and phone consistency across the top ten third-party profiles.
4. Twenty citable assets. Definitions, comparisons, cost tables, interval data, each with method, date and a named owner. This is the work that compounds.
5. The prompt panel. Fifty prompts, baseline run today, monthly cadence in the calendar.
6. Governance. A named owner for factual accuracy, an expiry date on every claim, and a quarterly re-run. Models change; your published prices change faster.

Why automotive is unusually exposed

Dealerships have the facts answer engines want and publish them in the worst possible form. Service pricing lives in a hero image. Availability lives in a widget. "Near me" questions are being answered in generated prose that names three stores, and the three are chosen on evidence the store never provided.

The store that states its service menu as data, its hours as data, its inventory as data, and its policies as data, consistently and with dates, is doing something its competitors are not. Not because it found a trick, but because it made itself quotable.

Bet on the invariants

The Wild West analogy is useful right up to the moment it becomes an excuse. The territory is unmapped, yes. But the things that survived 2003 are the same things working now: crawlable, specific, consistent, measurable, and owned by someone with a name.

Get ahead by being the store whose facts a model can trust. It is slower than buying a dashboard, and it is the only part of this that will still be true after the next model release.