13 rooftops in a week: what we found under the hood

We spent last week taking the Castle Automotive group from nothing to agent-native. 12 rooftops plus a flagship, every one migrated, rebuilt, measured and published.

12 of the 13 now score 100 out of 100 at Level 5, Agent-Native. The 13th sits at 93, and the reason it does is worth more to you than the 12 that passed.

This is the write-up. Not the case study version, the one with the parts that went wrong in it.

The robots.txt rule almost nobody applies correctly

Start with the thing that surprised us most, because it is probably live on your site right now.

A crawler reading robots.txt does not combine the rules that apply to it. It finds the most specific group matching its name and uses only that group. Everything under User-agent: * is then ignored completely.

John Mueller at Google confirmed the mechanism publicly this month, describing a case where a site had blocked its internal search results under the wildcard and Google kept indexing them anyway, because a Googlebot-specific section existed elsewhere in the file without the same rule.

Now apply that to what the industry has been doing for the past two years. Dealers have been adding AI crawler groups to robots.txt, usually pasted from a vendor blog post or a conference slide. The moment you add User-agent: GPTBot, that crawler stops reading your wildcard rules. Every Disallow you had carefully placed under * no longer applies to it.

We found this on our own output. Our generated robots.txt was granting AI agents access to paths the dealer had deliberately blocked for everyone else. On one store that included a lead form at /scheduletestdrive. We had not read the origin site’s rules and mirrored them into the agent groups, so we were quietly undoing the dealer’s own decisions.

The fix: parse what the site already disallows and inherit those rules into every agent group you write. It is now live across our fleet. But we shipped it wrong first, and if you have added AI crawler lines to your own robots.txt without doing this, it is wrong on your site too. That is worth checking this week regardless of whether you ever talk to us.

6 of the 13 could not be read at all

The second blocker was not ours and it was not the dealer’s.

6 of these rooftops run on a major dealer website platform. That platform answers AI assistants with a 403 and a challenge header. Not the dealer’s choice, not a setting in the dealer’s control panel, and in most cases not something anyone at the store knows is happening.

Worth being precise about the shape of it: the platform’s own published robots.txt contains no AI directives at all. The wildcard permits these assistants. So the crawl policy the platform publishes and the security layer it runs contradict each other, and the security layer wins.

The consequence is that a dealer can do everything right, buy every tool, publish perfect structured data, and still be invisible, because a bot rule three layers above them rejects the request before any of that is read.

If you are on a hosted dealer platform, the single highest-value question you can ask your vendor this quarter is whether Verified Bots is set to Allow in their edge configuration. It is one toggle. It costs nothing. Cloudflare’s verified list already validates GPTBot, ClaudeBot, PerplexityBot and Google-Extended by signature rather than trusting a user-agent string, so an impostor is still blocked.

Why the 13th store sits at 93

One domain is at 93, and we are publishing that number rather than the average.

It fails a single check: a DNS record used for agent discovery. We created the record correctly. It is right in every place we control. It is waiting on the registry side of an operation we do not run and cannot accelerate.

We could hide this. Report the 12, average the 13 to 99, put a badge on the site. We would rather tell you that our own scoreboard shows one of our stores below target, because a vendor whose dashboard never shows a problem is a vendor whose dashboard is not measuring anything.

That is not a principle we adopted for a blog post. Our score badge reads live from the same public checker you can run yourself, so when a store drops, our own marketing shows it dropping.

What the outside data says about why this matters

We are a vendor, so take our framing with the appropriate salt and check the sources.

SOCi’s 2026 Local Visibility Index looked at more than 350,000 locations across 2,751 multi-location brands. Two findings stand out. Business profile accuracy in AI answers ran at roughly 68% on ChatGPT and Perplexity, against 100% on Gemini, which is grounded directly in Google Maps. And there was only 45% overlap between the brands that rank well in traditional local search and the brands AI actually recommends. More than half of the businesses that are winning on Google are simply not in the answer.

The selectivity number is the one that should get your attention. ChatGPT recommended about 1.2% of the locations it was asked about. Google’s local pack shows roughly 35.9%. The list got about 30 times shorter, and nobody told the businesses that fell off it.

On accuracy, a study covered by Search Engine Journal this month tested 165 businesses across ChatGPT, Perplexity and Gemini roughly 13,000 times. 93% had at least one basic fact wrong or missing. Smaller businesses were hit hardest, with fabricated facts appearing at 50% for SMEs against 32% for large companies.

And on the buying side, Cox Automotive’s Car Buyer Journey study, fielded in fall 2025 across 2,300 recent buyers, found 19% of all buyers and 25% of new-vehicle buyers used AI tools while shopping. Those buyers were meaningfully more confident: 81% believed they got the best deal, against 67% of everyone else.

Then there is the number that changed our own thinking. Adobe Analytics tracked AI-referred traffic to retail sites and found it now converts about 42% better than traffic from everywhere else. A year earlier the same measurement had it converting at roughly half the rate. That is a complete reversal inside 12 months, on volume that grew several hundred percent year over year.

The early dismissal of AI traffic was reasonable at the time. People were experimenting, clicking through, buying nothing. That is no longer what the data shows. The visitor an assistant sends you today has already been qualified by the assistant.

That pairing is the whole argument. A growing group of shoppers is using assistants, they arrive readier to buy than any other channel, they trust the outcome more, and the assistants are wrong about local businesses most of the time.

One thing we will not tell you

There is a claim going around that publishing an llms.txt file makes you visible to ChatGPT. We publish llms.txt on every store we build, and it is genuinely useful, and that claim is not true.

Adoption sits around 8.7% of the top 1,000 sites as of June. More importantly, the major consumer assistants largely do not fetch llms.txt when answering a shopper’s question. It is read mostly by coding agents pointed at documentation. Anyone selling you an llms.txt file as the answer to “does ChatGPT know my hours” is selling you a real file that does not do that job.

The same goes for agents.md, which is a convention for software repositories, not a business data standard. Vendors are already conflating the two.

What does the job is duller: complete and correct structured data on every vehicle and every offer, a crawler policy that actually resolves the way you think it does, a live connection so an assistant reads current inventory rather than a stale snapshot, and measurement you did not write yourself.

Where the industry actually is

Every AI-visibility product we can find for dealers is a monitoring product. They tell you how you appear in AI answers and help you tune content in response. That is a real category and some of those tools are good.

It is also not the same job. Monitoring tells you that ChatGPT got your hours wrong. It does not make your hours machine-readable. We think the infrastructure has to exist before the dashboard means anything, which is why we build the layer and give the measurement away free at isitagentready.com.

Run your own store through it. If you land around 21 out of 100, you are in completely normal company, and now you know the size of the gap.

Welcome to the group

Castle Automotive joins Charlie Obaugh Auto Group, Bob Allen Motor Mall, Markquart Motors and Bud’s as stores whose inventory, hours, staff, departments and offers are now readable and queryable by any assistant that asks properly.

13 more rooftops that will not be missing from the answer.