Why ChatGPT doesn't recommend your business

Two things are usually true at once: the assistant cannot read your website the way it reads a directory's, and it does not know your business well enough to put a name behind. When we asked two AI assistants who to hire in eight American cities, 1,407 of 2,232 stored answers named a directory and six named the business the question was asked about. Here is what produces that, and what you can change this week.

What happens when someone asks an AI who to hire

Between 21 July and 11 August 2026 we asked the questions people ask when they are hiring, in nine American cities, and stored every answer in full. Plain questions: best wedding planner in the city, best wedding officiant, who should officiate my wedding. Each went to the API models behind ChatGPT (OpenAI's gpt-4o-mini) and Claude (Anthropic's Claude Haiku 4.5), three times per engine. Of the 5,904 stored answers, 2,232 answered a who-to-hire question.

A directory was named in 1,407 of them. The Knot appeared in 1,371, WeddingWire in 1,332.

Six answers named the business the question was asked about. Six, across eight cities.

That count is narrow on purpose. It does not mean the answers held no local names. It means that when we asked about a business's own city and trade, the business itself came back six times. From the other side: of 173 wedding businesses probed in Austin, Nashville and Charleston, 153 were never named in any stored answer on either engine.

Change the question from who to hire to where to get married and the directory rate collapses. Only 383 of 3,672 venue answers named a directory, about 10 percent, and real places filled the space: Biltmore Estate in 413 Asheville answers, City Park in 398 New Orleans answers. What those places have in common is size and fame.

So the assistant is not refusing to name specific businesses. Asked who, it does not know a person or a small firm well enough to stand behind a name, so it hands the customer to an intermediary.

Follow that customer one more step. A lead directory's business is selling introductions. When the assistant sends your customer there, the marketplace has them, and the way you get them back is to pay for them. If you already advertise on that directory, that is the arrangement working as designed: you buy an introduction the assistant just made for free, in the marketplace's favor rather than yours.

None of this is a verdict on the directories: they did the one thing most small business sites have not done, which is publish their information in a form a machine can read. It is also dated, and measured in wedding markets on two models, so read it as the shape of the problem rather than a forecast for your trade.

The three reasons our audits keep finding

Nothing on the page says who you are in a form a machine can read

A person reading your homepage knows in two seconds that you are a plumber who does emergency call-outs. A crawler gets prose, and prose has to be interpreted. The fix is a machine-readable identity block, JSON-LD, in the page source: business name, what you offer, where you work, how to reach you, as labeled fields rather than sentences. That is 25 percent of our readiness score, for a simple reason. When the facts are labeled, an assistant quotes them. When they are not, it guesses, and the guess lands on whatever it is surer about.

The crawlers cannot fetch your pages at all

This is more common than it sounds, and almost never a decision anybody made. On 16 July 2026 a working, well built site scored 60 out of 100 on our audit. The next day, with no copy rewritten and no page redesigned, it scored 93, and 93 again two weeks later. The site is vonzie.studio, the agency site of Vonzie Studio, the studio that runs The Answer Report, which is why we can publish the before and after audits in full.

Two things were wrong. Its robots.txt, the file that tells automated readers what they may look at, was telling GPTBot, ClaudeBot, Google-Extended, CCBot, Bytespider, Amazonbot and Applebot-Extended to leave. The owner had written none of it. It was Cloudflare's managed robots.txt, applied by the platform to keep content out of AI training sets. Second, the audit read five pages and found no phone, no email and no address, while a visitor saw the email plainly. That was Cloudflare's email obfuscation, which rewrites addresses into a script the browser reassembles. AI crawlers do not run scripts, so the address was on the page and invisible at once.

Neither setting is a bug. The trouble is that the switch keeping your content out of training sets also keeps your business out of the answers, and nothing tells the owner which game they just opted into. In our 31 July sweep, two more Cloudflare-fronted sites were serving the identical managed file, line for line. The pattern travels with the platform, not the site. A third case belongs here too: text that only appears once JavaScript runs, where a no-script reader gets an empty shell.

You are not present where the answers already read

The answers above are full of directory names, review platforms and well-known places. That is where these systems find dense, consistent, checkable information about local businesses. If your listings are thin, stale, or carry three versions of your name, address and phone number, there is nothing to resolve into a confident recommendation. We do not score that half today, and we would rather say so than let you assume it counts.

What to do this week, in order

1. Open your own domain followed by /robots.txt in a browser. If you see AI crawler names each followed by Disallow, that is the first fix. On Cloudflare the setting is managed robots.txt, and email obfuscation sits under Scrape Shield. Both can be switched off, or replaced with rules you chose deliberately.
2. Put your phone number, email address and street address in the page as plain text and ordinary links, on the homepage and the contact page. No image of a phone number, no script that assembles the email in the browser.
3. View the page source and search it for a sentence you can see on screen. If it is not there, your text arrives via JavaScript and a crawler is reading an empty shell. That one goes to whoever built the site.
4. Add a JSON-LD block naming the business, its category, its service area, its address, its phone number and its hours. Most platforms have a way to do this, and it takes a web person under an hour.
5. Make your name, address and phone number identical everywhere they appear, on your site and on every listing you have. Pick one form and stop varying it.
6. Publish a sitemap, and add an llms.txt file pointing an AI reader at the pages that describe what you do.

That order matters: there is no point labeling your facts for a crawler that has been told to stay out.

How to know whether it worked

Two checks, answering different questions.

Ask the way a customer asks. Open a fresh chat and type your trade and your city, not your business name. Ask both ChatGPT and Claude, ask each more than once, and write down what came back, with the date. One answer is noise. A pattern across several runs on two assistants is a reading.

Then check the machine-readable half, the half you control and the half that moves in days rather than months. The free audit on our front page fetches your site the way a no-script AI crawler does and quotes the exact line behind every deduction, so you can hand the list to your web person and run it again next week.