A customer found Cartage by asking ChatGPT. So I did the same work on 11 sites.

The test: 39 questions before changing anything

To see where everything stood, I wrote three or four questions per site, worded as a customer would type them and never naming the product. For example: "Is there a Mac app that shows Finder file copy progress in the menu bar?" Each question went once to OpenAI's gpt-5-mini model with its web search tool switched on, which is the closest thing to ChatGPT's search that can be run from a script. I recorded whether the answer or its cited sources named my product or domain.

Some caveats, because a table looks more scientific than it is. Each question was asked once, so a single yes or no is a sample, not a measurement. It covers one assistant. And the sites differ in age and competition, so preparation is not the only reason for the gap.

Still, the pattern was hard to ignore. The three sites with real content and machine-readable facts were the three that got named. Where my sites lost, it mostly cited small competitors' own websites, far more often than the App Store. One weather question was answered with programming code and no sources at all: the assistant did not know any such app existed.

What I did to all 11 sites

On 10 November 2026, in one pass, every site got:

• an /llms.txt, and an /llms-full.txt where there was enough content to justify one;
• a robots.txt that names the AI crawlers and Bing's crawler and lets them in;
• valid structured data: the product, its price where one is public, the company, and the FAQ;
• a canonical address on every page;
• an IndexNow key, with every page submitted so search engines knew about the changes at once;
• new content where a question had no answer, such as a comparison guide for Reroute and a guide for ViewPortal.

Afterwards, an automated check confirmed that on all 11 sites the crawler rules, llms.txt, sitemap, structured data and IndexNow key were live and valid, and that requests identifying as GPTBot and ClaudeBot got a normal response. Each home page scored 100 for SEO in Google's Lighthouse test.

What I found that mattered more than the clever bits

This is the part I would want to know if I were hiring someone. Going through each site as an assistant would turned up plain, embarrassing faults that no amount of structured data would have fixed:

• A download button that had never worked. On Candle & Dagger's site, a coding slip meant the App Store button pointed nowhere. Visitors saw "Download on the App Store" and could not.
• Two app sites with no link to the App Store at all. An assistant that found FourCast or Back Seat Video Remote had nothing to send people to.
• Three pages describing the wrong product. Iconic Icons' pricing, gallery and developer pages still carried a description left over from a template.
• Three sites answering "yes" to a file that did not exist. Asked for /llms.txt, they returned their home page with a success code. A machine cannot tell that from a real file without reading it.
• A security service quietly rewriting robots.txt. On two domains, Cloudflare was adding rules that told GPTBot, ClaudeBot and others to stay out. I had not asked it to.

An assistant can only recommend what it can reach and make sense of. Most of the value was in finding the things that were simply broken.

What went wrong

Two things, both mine. One site's home page returned an error for about fifteen seconds after a deploy, because a file was uploaded with the wrong permissions. I put it right at once and changed the deploy script so it cannot happen again. And one deploy failed before uploading anything, thanks to a typing mistake, and had to be re-run. Nobody noticed either, but you should know how I work when things go sideways: I check, I fix, I write it down.

What it cost to measure

The 39 questions cost about $1.29 in fees to OpenAI, by a deliberately cautious estimate. I mention it because "AI visibility tracking" is often sold as something exotic. The asking is cheap. The value is in choosing the right questions and knowing what to do with the answers.

The result, so far

Here is the honest position on 11 November 2026. The work is live and verified on all 11 sites. Whether ChatGPT now names more of them, I do not know yet. Assistants and search engines have to crawl the sites again, and I allowed two to three weeks for that before re-running the same 39 questions.

When the re-run is done, the numbers will go in the table above, good or bad. If the answer is "it moved three sites and did nothing for the rest", that is what this page will say.