How to get recommended by AI

Why does AI ignore a site that ranks well on Google?

Because they are different jobs. A search engine returns a list and lets you choose. An answer engine retrieves passages from a handful of sources, weighs them, and writes one answer that names two or three brands. Ranking tenth on a list still puts you on the page. Coming fourth in an engine's internal ordering puts you nowhere.

There is a second reason, and it catches people out more often. AI recommends brands through two separate channels. Memory is what the model absorbed in training: instant, sourceless answers that move on training timescales and are shaped by what the wider web says about you. Retrieval is a live search at answer time, with citations, shaped by whether your pages can be reached and parsed right now. Work that moves one may not touch the other.

How rare is a site that is actually ready?

Rare enough to be the whole opportunity. Across 1,000,000+ analysed websites, 216 score AI-Ready, which is 0.022% of them, or about 1 in 4,600. The highest score recorded anywhere is 86.6 out of 100, so nobody has reached the top band.

Source: SearchScore's corpus of 1,000,000+ analysed websites, current as of 2026-07-27.

Step 1: measure before you change anything

Write down the questions a buyer would actually type before they know your name. Not "who is [your brand]", which you will always win, but "best accountant in Manchester for a small business" or "software for signing contracts securely". Twenty to thirty of them.

Ask each one across the engines that matter to you, and do it more than once. This is the step people skip and it is the one that makes everything after it meaningful. Research published in 2026 measured the source overlap between two same-day runs of the same prompt at roughly a third, so a single run tells you almost nothing. Record how often you are named, as a percentage, and keep the questions locked so next month's number is comparable.

Step 2: foundation, so a machine can read you at all

Three jobs, mostly one-off.

Let the right crawlers in. Each AI company runs separate bots for separate purposes. Search bots, such as OAI-SearchBot and PerplexityBot, govern whether you can be cited. Training bots, such as GPTBot, govern what the model remembers about you. Google's AI surfaces are fed by ordinary Googlebot. Blocking the wrong one is a common and completely silent cause of invisibility, and a firewall or CDN rule can be doing it without anything in your robots.txt saying so.

Say what you are. An engine can read every word on your site and still not be confident what category you are in or when to recommend you. Structured data, a consistent description across every profile you own, and a Wikidata entry all reduce that ambiguity. Treat schema as disambiguation rather than as a ranking trick: Microsoft confirms it uses structured data for Bing and Copilot, Google does not require it for AI features, and no controlled study shows it buys citations on its own.

Be consistent. Name, description and category identical everywhere. Machines cannot confidently merge two spellings of your company into one entity, and an entity split in half is an entity with half the evidence.

Step 3: extraction, so your answer can be lifted

Retrieval works on chunks. A model pulls a passage out of your page and uses it, so the passage has to make sense with nothing around it.

Lead with the answer. Traditional copy sets the scene and delivers the payoff at the end, which is exactly wrong here: put a direct, self-contained answer in the first two sentences under a heading, then support it. Use headings that match the question a person would ask. Keep the answer paragraph short enough to be quoted whole.

Then cover the decision rather than perfecting one page. Engines split a question into many hidden sub-queries and retrieve for each, so a business that answers eight related questions on eight clear pages beats one that answers all eight halfway down a single long one.

Step 4: reinforcement, because other people decide

This is the layer most businesses skip and it is the one that decides evaluative answers. When an engine is asked who is best, it leans heavily on what sources other than you say. Directories, review platforms, trade press, forums and comparison pages carry more weight in that judgement than your own claims about yourself.

The reliable version of this is not "earn mentions", which is advice the way "be rich" is a financial plan. It is a specific list: claim and complete the profiles that already rank for your category, get reviews onto the platforms your buyers read, and publish something worth citing.

The most effective single move available to a small business is to own a number. Publish original data about your own niche, however small the sample, and describe the method honestly. A number with a method attached is the thing writers cite, and citations are what engines read.

Step 5: track it, and repair what goes wrong

Re-run the same locked questions on a schedule and watch the rate. Scatter between runs is noise. A slope over months is signal. Two objective measurements are worth adding alongside: your own server logs, which show which AI crawlers actually fetch your pages, and your analytics, which show sessions referred from assistants. Referred sessions are a floor rather than a measure of influence, because most answers are read without anyone clicking.

Sooner or later tracking will surface something worse than absence, which is an answer that names you and gets you wrong. Wrong prices, a service you stopped offering, a feature you do not have. That is a correction job with its own sequence, and it starts with finding the source the engine is reading.

What order should you do this in?

Foundation, then Extraction, then Reinforcement. The order matters because each layer depends on the one before it. Earning a mention on a page an engine cannot crawl buys you nothing, and a beautifully structured page nobody corroborates will get you considered and not chosen.

Common questions

How long does it take to get recommended by AI?

Retrieval-layer work shows up fastest. Crawler access, structure and schema can change what an engine is able to cite within weeks, because the engine is reading your live pages at answer time. Memory-layer work is slower, because it depends on what the wider web says about you and on model updates. A realistic expectation is movement on retrieval-driven engines inside a quarter, and slower movement on the sourceless answers.

Does llms.txt get you cited?

No major answer engine uses llms.txt as a citation signal, and Google has said so on the record. It takes five minutes and is worth doing as insurance for the agentic web, and it should not be treated as a lever.

Does schema markup get you cited?

Microsoft confirms it uses structured data for Bing and Copilot. Google does not require it for its AI features, and no controlled study shows it buys citations directly. Add it for disambiguation and entity clarity, which is a real benefit, rather than as a ranking trick.

Can I just do this once?

The Foundation work is largely one-off. The measurement is not. Same-day runs of the same prompt overlap by roughly a third, so visibility has to be read as a rate across repeated runs, and it drifts as models update and competitors improve.

Is this different from SEO?

It overlaps heavily and the goal is different. SEO competes for a position in a list of links. This competes to be one of the few sources an engine uses when it writes a single answer. A site can rank first on Google and never be named by ChatGPT, which is the situation this method exists to fix.