How to improve AI visibility

Most brands that are missing from AI answers assume it is a content problem on their own site. Usually it is two problems at once: assistants cannot read the site cleanly, and nobody outside the brand's own domain describes what it does. Fixing only the first rarely changes the answers.

The order below matters. Steps 2–4 are cheap and mechanical. Steps 5–6 are where the compounding happens.

The nine steps

1. Measure a baseline. Run a fixed prompt set across ChatGPT, Gemini, Claude, Perplexity and Copilot and record mention rate, competitor share of voice and the URLs each model cites.
2. Let AI crawlers in. Allow GPTBot, ClaudeBot, PerplexityBot, Google-Extended and Bingbot in robots.txt, keep private areas disallowed, and confirm your pages return real HTML rather than an empty shell.
3. Publish llms.txt and a clean sitemap. Give models a plain-language map of your most important pages and keep the sitemap returning HTTP 200 with no stale URLs.
4. Add structured data. Organization, Product, Offer, FAQ and BreadcrumbList schema turn marketing prose into facts a model can quote without inferring.
5. Write the page types models quote. Definitions, comparisons, category roundups, pricing explanations and direct answers to real buyer questions — each answering one question clearly near the top of the page.
6. Earn third-party mentions. Get listed on review platforms and category directories, appear in roundups, and take part in genuine community discussion. Models weight corroboration from sources you do not own.
7. Make your facts consistent. Align naming, pricing, category language and claims across your site, listings and profiles so a model is not choosing between contradictory sources.
8. Become transactable. For commerce, publish a product feed, structured availability and pricing, and an agent endpoint (MCP or UCP) so assistants can act on your catalogue rather than just describe it.
9. Re-measure on a schedule. Run the same prompt set weekly. Answers are non-deterministic, so only a repeated sample separates a real change from noise.

Why third-party mentions dominate

When an assistant recommends vendors in a category, it rarely quotes a vendor's own homepage — it quotes the pages that compare vendors. Roundups, review platforms, directories, analyst notes and forum threads are the corroboration models lean on, because they read as independent.

• Get listed on the review platforms buyers in your category actually use.
• Submit to credible directories in your niche rather than link farms.
• Pitch to publish in existing roundups — being added to a "best X" article that already ranks beats writing a new one from scratch.
• Take part in real discussion where your category is debated. Models retrieve community threads heavily.
• Make partner, integration and customer pages on other sites name you explicitly.

What to change on your own site

Structure over style

Lead each page with a direct one-paragraph answer to the question it targets, then support it. Models extract the first clear, self-contained statement they find. Long scene-setting introductions get skipped.

Facts as data

Prices, specifications, availability, delivery times and return policies should exist as structured data, not only as prose or PDFs. If a model has to infer a fact, it will often decline to state it.

From described to usable

The end state for commerce is not just being mentioned. It is an assistant being able to search your catalogue, read live pricing and hand a customer to checkout — which is what an agent endpoint and protocol readiness deliver.

What to expect

Retrieval-based surfaces such as Perplexity and browsing modes can reflect new pages and mentions within days. Answers that lean on training data move over months. Track the trend, not any single answer — the same prompt asked twice will not give you the same words.