How to improve your business's AI visibility: 7 steps in order

To improve your business's AI visibility, you need to get three things right, in this order: the crawlers behind ChatGPT, Claude, Gemini and Perplexity must be able to read your site, every public source must say the same thing about you, and independent third parties must name you. The order matters, because an AI that cannot read your pages will not cite your answers, however well written they are. Seven steps cover it all: check where you stand, open the door to crawlers, align your sources, answer first on every page, show your proof, get named somewhere other than your own site, then measure over time. None of them involves buying a placement: paid content accounted for 0.3% of the sources AI cited in the largest study published on the subject.

On this page

This article is for SMB owners and marketing teams, B2B and B2C alike, who wonder why an AI names their competitors and not them. It gives the overall plan. Each step links to the analysis where we go deep on the subject, with the method and the sources.

Why AI visibility is worth the effort

Your customers already put their buying questions to assistants. In a Gartner survey of 646 buyers conducted from August to September 2025, 45% of B2B buyers said they had used artificial intelligence during a recent purchase. On the consumer side, Caila Schwartz of Salesforce, quoted by Retail Insider on September 1, 2026, estimated that about 11 to 12% of Canadian shoppers turned first to AI tools such as ChatGPT, Gemini or Perplexity.

The volume of visits is still modest, but those visits are valuable. In May 2026, Adobe measured AI traffic to US retail sites converting 54% better than other channels, while noting it remained small compared with paid search or email. The detail is in our analysis does traffic from AI convert better?

Then there is what no report shows. When a buyer asks an assistant to narrow a market down to three names, the business left off the list does not lose a sale: it is never invited to bid. We describe that mechanism in how B2B suppliers get recommended to buyers by AI.

One caveat: these figures come from the United States or from surveys. They give a direction, not a forecast for your market.

How an AI chooses the businesses it names

An assistant does two things when it gets a buying question. It draws on what it learned during training, then, more and more often, it runs a web search at answer time and reads a few pages to back up its answer. That second step is the one you can influence quickly.

Each assistant searches in its own way. Measurements published in the summer of 2026 showed ChatGPT querying OpenAI's own index first, then Google results retrieved by contractors. Claude goes through Brave Search, which is on Anthropic's public list of subprocessors. Gemini and Google's AI Mode rely on Google's index. Ranking first on Google does not automatically put you in what ChatGPT or Claude see of you. We documented these paths in what search engine does ChatGPT use? and in Claude and Brave Search.

What these systems have in common matters more than their differences. All of them must be able to read your page, understand what you do and for whom, then find confirmation elsewhere. Hence the order of the seven steps.

The 7 steps, in order

1. Check where you stand, over several runs

Ask assistants the questions your customers ask, first without naming your business, then naming it. Do it on at least two assistants, three times, a few days apart. Note three things: whether you are named, whether the description is accurate, and which sources the assistant cites.

A simple benchmark is enough to track the result over time: the number of answers that name you, divided by the number of questions asked without your name. Four answers out of ten gives 40%. It is a home-made benchmark, not an official ranking, and it only holds if you ask the same questions the same way each time.

Repetition is not a luxury. Semrush and Kevin Indig put 100 buying questions to ChatGPT in two of its modes: only 25.6% of cited domains overlapped from one mode to the other. A single answer, good or bad, tells you almost nothing. Judge by what keeps coming back.

The list of cited sources is the most useful output of the exercise: those are the pages where you will need to appear in step six. The full method, in one hour, is in does AI understand your business?

2. Open the door to AI crawlers

This is the cheapest step and the one most often neglected. Three checks are enough.

Does your text arrive without JavaScript? In December 2024, Vercel and MERJ analyzed AI crawler behaviour across millions of requests: none of the crawlers from OpenAI, Anthropic, Meta or Perplexity executed JavaScript. Only Google's and Apple's rendered pages the way a browser does. The study is two years old and crawlers may have changed, but the caution stands: the main text, the headings and the prices should be in the page's HTML. To check, view a page's source code in your browser: your text should be readable there in plain form.

Does your robots.txt file let the right crawlers in? Two families visit your site. Training crawlers collect content to feed a model. Search crawlers read a page at the moment someone asks a question, then cite the source. Blocking them together is the most common mistake. OpenAI writes that a site that blocks OAI-SearchBot does not appear in ChatGPT's search answers. You can refuse training and stay visible: the split is explained in AI crawlers: which should you allow or block?

Do you have a rule that lets only Google in? Brave's crawler does not announce itself under a distinct name and follows what your site grants Googlebot. A rule that blocks everything except Google can remove you from what Claude sees, with nothing to tell you so.

To go further, your server logs show which pages AI systems actually read. The recipe, with a simple spreadsheet, is in AI crawler log analysis.

3. Say the same thing everywhere

An AI does not read only your site. It cross-checks your site against your Google Business Profile, your LinkedIn profiles, your industry directories, the press and even your job postings. When those sources contradict each other on what you do, for whom or in which region, the description gets blurry and the recommendation can go to the wrong type of customer.

The sources most often in contradiction are the ones nobody watches: an old directory listing, a LinkedIn profile that describes the company of five years ago, a job posting that advertises a service you no longer offer. Fix your site first, then your Google Business Profile, then profiles and directories, because some sources feed others.

The Google Business Profile deserves separate attention. An analysis published by Profound in July 2026 showed google.com had moved to second place among the domains cited by Google's AI Mode, mostly because of business profiles, in the sectors where decisions are made locally. The analysis comes from a tool vendor that does not detail its method: we read it as a signal. The settings that matter are in Google Business Profile: how to improve your local visibility.

If an AI is already saying something false about you, there is no button to correct it. The only route is to correct the sources it reads: see what to do when AI gets your business information wrong.

4. Answer first, explain after

An AI picks up passages, not pages. The Pew Research Center measured that a summary generated by Google ran to a median of 67 words and that 88% of those summaries cited at least three sources. Two consequences: your answer has to fit in one short, complete paragraph, and there is more than one spot to win.

In practice, every service page and every article opens with the answer to the question its title asks, before any background. Section headings use your customers' questions, in their words. An FAQ answers what people really ask you on the phone, never a question nobody asks. The method is detailed in the answer-first approach.

Three points are missing from most guides.

• Text has to be text. A spec sheet in a PDF or features inside an image drop out of answers. Everything a buyer needs to choose you should be written on the page.
• French is a condition in Quebec. A customer who asks in French gets sources in French. A catalogue that exists only in English drops out of those answers.
• Structured data helps understanding, it does not earn citations. Schema.org markup tells machines who you are, where you are and what you sell. Google writes, however, that no special optimization is required to appear in its generated answers. Markup adds to good content, it does not replace it.

5. Show what makes you recommendable

To recommend one business over another, an assistant needs a reason it can read. Five elements provide it.

• One page per service. Each important service has its own page, which answers five questions: who you are, what exactly you do, for which type of customer, in which region, and why you. A single page that lists all your services answers no specific question.
• A named specialty. Google writes that people ask longer and more specific questions in its AI search experiences. "Accountant for dental clinics in Laval" answers those questions; "quality accounting services" does not.
• Published prices. A page that gives your starting prices and what they include answers a question every buyer asks. Without it, the assistant cites a competitor's price or a comparison site's.
• Named people. Articles signed by an expert, with their role and profile, give the AI a source it can cross-check.
• Concrete cases. A named client, their problem and the approach, in text, with their written consent. It is the only proof a competitor cannot copy.

Freshness counts too. In Muck Rack's February 2026 study, more than half of the citations observed had been published in the previous twelve months. A page from 2022 that nobody has touched starts at a disadvantage. Date what has actually changed, without dressing up an old page as a new one.

6. Get named by third parties

This is the step that weighs the most and the one guides cover the fastest. Three studies let us be precise.

Comparisons. Peec AI classified 1,056,727 citations collected from 75,000 answers by ChatGPT, Google's AI Mode and Perplexity. On commercial queries, 40.86% of cited sources were comparison listicles. In professional services, 80.9% of the most cited listicles were written by third parties rather than by the companies that appear in them. Publishing your own ranking with yourself in first place therefore carries little weight. You need to find the comparisons that already exist in your category (step one gave you the list), understand their criteria, then do the legwork to get in.

Mentions. Zyppy evaluated 23 factors in May 2026: the correlation between a brand's mentions on the web and its citations by AI was 0.664, against 0.218 for inbound links. Being named, even without a link, counts about three times more than getting a link.

The press. In May 2026, Muck Rack analyzed more than 25 million links cited by ChatGPT, Claude and Gemini: 27% came from journalism, against 0.3% for paid content and advertorials. An article in a trade publication in your sector is worth far more than sponsored copy.

Two sources close to you complete the picture. In a Semrush study of 325,000 queries, LinkedIn was the second most cited domain. Among the LinkedIn content cited, 59% came from individual profiles rather than company pages. The experts on your team who publish under their own name weigh as much as the company page. Your customers' reviews, for their part, can be summarized by the assistant before it recommends: their regularity and your replies count as much as the rating. A useful review tells a story: the starting problem, the service chosen, the result, the region. Invite your customers to describe their experience rather than just leave stars.

The steps to follow are in getting cited by AI: the role of comparisons and brand mentions.

7. Measure over time, with first-hand sources

Three free sources measure your visibility at the source rather than by replaying questions.

• Bing Webmaster Tools has offered a report on citations in Copilot since February 2026, with the searches the assistant runs to build its answer. It is the most useful data of the lot, even if most of your traffic comes from Google.
• Google Search Console has had a generative AI performance report since June 2026: impressions by page, country and device, with no queries and no clicks.
• Your server logs show which pages ChatGPT, Claude and Perplexity came to read. It is the only source that belongs to you.

None of them sees everything: Bing sees Copilot, Google sees Google and neither sees ChatGPT. So redo the test from step one once a month, with the same questions. A paid tool becomes useful when you need to track several assistants in one dashboard, month after month. The selection criteria are in measuring AI visibility: which tools should you use?

Be careful with promises of results. Nobody today knows how to attribute a sale to a citation in an AI. What can be measured honestly is your presence in answers, the accuracy of the description and the share of citations you earn against your competitors.

What Google's AI Mode and ChatGPT answer, and what they leave out

On November 5, 2026, we put the same question, "how do I improve AI visibility for my business?", to Google's AI Mode and to ChatGPT, in French, from Quebec.

Google's AI Mode answered with four pieces of advice. They are sound. The table shows what the data adds to each.

ChatGPT, for its part, produced a guide in thirteen sections: the website, crawler access, external platforms, content, structured data, reputation, testing, then a 90-day plan. All of it is sensible and we borrow two ideas from it above, the bank of test questions and the page per service. Its answer, however, cited no figure and no study, said nothing about JavaScript, did not name Claude among the assistants to watch and proposed a budget split in percentages without saying where it came from.

One detail of Google's answer says more than its advice. Among the sources it displayed were agency articles, but also a LinkedIn post signed by an individual, a YouTube video and a Reddit thread. The pages that feed an answer are not all company websites: that is step six, seen live.

What both answers lack most is order and sorting. They put steps of very different weight on the same footing and say little about what can be left aside.

What is a waste of effort, or nearly

The llms.txt file. This file, proposed in 2024, tells models which pages matter on your site. In the fall of 2026, no major engine said it read it. Google also writes that no machine-readable file is needed to appear in its answers. The exercise of writing it has value, because it forces you to choose your important pages. The file itself will not get you cited. Our analysis: does llms.txt help your visibility in AI answers?

Advertorials and mass-distributed press releases. In Muck Rack's studies, paid content accounted for 0.3% of citations and press release distribution sites for 1%.

Advertising in ChatGPT. Paying for an ad does not make you a source of the answer. We tested it in B2B in Canada in the spring of 2026: a 0.9% conversion rate against 5.95% on Google Ads, for the same client over the same period. The write-up is in ChatGPT ads in Canada: results from our B2B test.

Blocking Google-Extended out of fear, or allowing it out of hope. This setting affects the training of Gemini models, not Google Search. Blocking it does not remove you from Google and allowing it does not move you up. The detail: does blocking Google-Extended affect your Google rankings?

Bought reviews and links. An assistant cross-checks its sources. Fake reviews or hundreds of purchased links add no independent confirmation. Fake reviews also expose you to penalties from the platforms.

Copy written "for AI". Pages stuffed with keywords or mass-produced by an AI bring nothing an assistant cannot already find elsewhere. Google advises content that is not interchangeable with everyone else's: your figures, your cases, your experience.

Where to start, depending on your situation

In every case, the first two steps take a few days and cost next to nothing. The sixth takes months. One more reason to start early.

A 90-day plan

The seven steps fit well into three months. Durations vary with the size of your site and the number of sources to correct.

Three months are enough to lay the foundations and get a first comparison. They are not enough for step six, which carries on afterwards. No calendar guarantees a recommendation: it gives an order of work.

How this ties in with your SEO

AI visibility does not replace search engine optimization, it builds on it. Google writes that the same best practices apply to its classic search and to its generated answers. A fast, well-structured site with useful content has a head start in both.

Three workstreams are added: measurement, which runs through other sources, the consistency of what is said about you off your site, and mentions by third parties. We describe them in SEO vs. GEO: what's the difference? If the acronyms lose you, GEO, AEO, AIO, SXO and SGE untangles them.

How would you like to send this email?

Gmail ↗Outlook ↗Email app on this computerCopy the message

Close

Frequently asked questions