Teckel AI · AI discovery guide

How to get ChatGPT to recommend your business

Updated: September 2, 2026 · By Teckel AI

You work two surfaces at the same time. On your own site, the essential content has to be extractable from the HTML, each page has to answer the customer's question in its first two paragraphs, and your entity details have to be consistent everywhere. Off your site, you have to be named in the third-party sources the engines already cite for your query.

Most guides on this topic only cover the first surface. The citation data says that is the smaller half of the job. Muck Rack's May 2026 edition of its Generative Pulse study analyzed more than 25 million links from ChatGPT, Claude and Gemini responses across 17 industries and found that 84% of AI citations are earned media, meaning journalism, academic and government sources, encyclopedic sites and third-party corporate content, while paid and advertorial content accounts for 0.3%. Journalism alone is 27% of cited sources. None of this guarantees a citation or a recommendation. It tells you where the leverage is.

Why is your own website the smaller half of the answer?

Because in a commercial query, the thing the engine cites is usually not a company's website. It is a list that contains that company. Wix Studio's 2026 analysis of more than a million citations found that listicles capture 40.9% of commercial citations. When a buyer asks an engine for the best vendor in a category, the engine leans on somebody else's roundup.

Greg Galant, co-founder and CEO of Muck Rack, put the consequence plainly when the May 2026 numbers came out: "If your brand is not showing up in the media coverage AI is reading, you are not showing up in the answers AI is giving." That is a blunt way of saying that a strategy made only of publishing on your own domain has a ceiling built into it.

So run this diagnostic before you rewrite anything. Open the engine your buyers use, ask your highest intent commercial question, and read the list of sources the answer cites. Then sort those sources into three piles: lists you could plausibly be added to, directories you can register in, and publications that would have to cover you. That list is your actual work queue. Rewriting your homepage a fourth time is not.

What has to be true before an answer engine can use your page at all?

Four conditions. Each one you fail is a reason for the engine to use somebody else's page instead.

How do you write a page an answer engine can quote?

The academic work here is unusually specific. The GEO paper from Princeton, Georgia Tech, the Allen Institute and IIT Delhi, presented at KDD 2024, tested content changes against a generative engine and measured which ones moved visibility. Two results stand out: adding quotations from relevant experts lifted visibility by up to 41.8%, and adding statistics with cited sources lifted it by up to 32.8%. Keyword stuffing, the tactic borrowed from the old playbook, did not help.

• Lead with the answer. Before the story, before the credentials, say what you do, for whom, and what it costs. A reader who bounces is a lost visit; an engine that cannot locate the answer moves to the next source.
• Use real questions as headings. Not clever section titles. The literal phrasing a buyer would type.
• Quote named experts and cite your numbers. This is the highest measured effect in the literature and it is also just better writing. Attribute every statistic to a named source with a date so a reader can check it.
• Publish what your competitors hide. Prices, timelines, and the cases where you are the wrong choice. Specificity is extractable; adjectives are not.
• Mirror your FAQ to your schema. Structured data should describe content that is visible on the page, never content that only exists in the markup.

What about llms.txt, schema, and the rest of the checklist?

Keep them in proportion. Valid structured data helps describe your content and costs little, but it does not guarantee ranking, citation or recommendation. As for llms.txt, John Mueller of Google Search Relations wrote on Bluesky in June 2025: "FWIW no AI system currently uses llms.txt." Publish one if you want a clean summary of your own canonical facts. Do not let it displace an hour of work that has measurable effect.

The same discipline applies to the traffic numbers that get quoted in this space. Imperva reported in 2026 that automated traffic passed 53% of all web traffic, and Cloudflare measured in July 2025 that some AI crawlers fetch enormous numbers of pages per referral they send back, with ratios reported at 1,091 to 1 and 38,065 to 1 depending on the crawler. Those are industry aggregates. They describe the environment, not your site. What Semrush's 2025 data does suggest about the visits that do arrive is that they convert at about 4.4 times the rate of traditional organic traffic, which is why the small numbers are still worth the work.

What do you do if another company shares your name?

This is the section most guides skip, and it is the one that silently wastes the most effort. If another organization shares your brand name, the engine's problem is not that you are uninteresting. It is that it cannot tell which entity you are. Publishing more pages does not fix an identity problem, and you can spend months on content while the collision quietly eats every result.

You can detect it in five minutes: ask each engine your own brand name in a clean session and read what comes back. If you get the other company, a same-name app, or unrelated results, you have a collision rather than a content gap. What actually helps:

• One canonical identity everywhere. The same legal name, category, city and contact details on your site, your business profile and every external directory. Inconsistency is what makes an entity ambiguous.
• Organization structured data with sameAs. Point explicitly at the external profiles you control so the graph has edges connecting them to your domain.
• A knowledge base entry. Encyclopedic sources are part of the 84% earned media surface, and they are the reference layer that disambiguates entities by design.
• An unambiguous about page. State what the company is, where it operates, when it was founded, and what it is not. If the collision is well known, say so in plain language.
• Third-party mentions that link back. One external source that names you and links your domain does more for disambiguation than ten pages you publish about yourself.

Treat it as a tracked metric, not a one-time fix. Measure your own brand query on the same cadence as your commercial queries, because entity resolution moves in both directions and you can lose ground you had.

How do you know whether any of this is working?

Two measurements, and the discipline to keep them separate.

• Ask the engines directly, on a schedule. Run the same customer questions in clean sessions across ChatGPT, Claude, Perplexity and Gemini. Record four things: whether you are mentioned, whether you are cited with a URL, which competitors appear, and which entity you get confused with. Repeat weekly with identical prompts. One engine is not representative of the others, and a single run is not a trend.
• Read your server logs, not just analytics. Logs capture requests that browser-based tools miss. Keep crawler hits, mentions, citations, referrals and closed business as five separate numbers. A crawler hit does not prove a person saw a recommendation, and a user agent can be spoofed.

The number that matters at the end of the chain is not visibility. It is whether a referral became a session, a lead, an opportunity and an invoice inside your own system. Everything before that is a leading indicator.

A six week plan you can actually run

• Week 1. Baseline. Run your top ten customer questions across four engines and save the answers with dates. Note every domain the engines cite. That domain list is your outreach map.
• Week 2. Extraction. Check the raw HTML of your three most important pages. If the content is not in the response, that is the first fix and it comes before everything else.
• Week 3. Entity. Make the name, category and contact details identical across your site, your business profile and every external profile. Add Organization structured data with sameAs. Run the brand collision test.
• Week 4. Rewrite two pages. Answer first, questions as headings, a named expert quote, statistics with sources, and your price if you publish one.
• Week 5. Off-site. Take the domain list from week 1 and pitch the lists and directories you could plausibly join. This is the 84% surface and it is the week most companies skip.
• Week 6. Re-measure. Same prompts, same conditions, compared against the week 1 baseline. Report what moved and what did not, without converting it into a promise.

Where Teckel AI fits

Teckel AI is an applied artificial intelligence consultancy. This site runs part of what this guide describes: every page is served as static HTML with structured data, and we log server-side requests associated with declared bot user agents. Those hits can be spoofed and they do not prove a citation, a recommendation, a human visit or a closed deal, which is exactly why we keep them as a separate number from the rest.

If you want this run on your business, the entry point is a free diagnosis: you tell us your case and we tell you where you stand today in front of the engines and what to move first. A deeper strategy report is $50 USD. If the problem turns out to be the website itself, our AI Websites start at $2,500 USD, and a full Implementation Sprint starts at $5,000 USD with a demo in 48 hours and production in 21 days.

Frequently asked questions

Go deeper: what an agent-ready website actually is, who builds AI-readable websites, and the Spanish edition with the Mexican market detail: cómo hacer que ChatGPT recomiende tu negocio.