If an AI answer recommends three companies and yours is missing, publishing more pages is an expensive first guess. The obstacle might be access, an unanswered buyer question, weak evidence or simply variation between answers.
Start by finding which of those problems you actually have. This guide proposes a 30-day workflow for a small team, with a record of what changed and a way to check the result. The schedule is a working plan, not a promise of visibility within 30 days.
On this page
- 01Decide what visibility means for your business
- 02Remove access problems before rewriting content
- 03Give each important question a useful destination
- 04Make your identity and evidence easy to verify
- 05Build a measurement workflow with free tools
- 06Run the work in four weekly passes
- 07Know what the result cannot prove
Decide what visibility means for your business
A brand mention, a citation and a visit are different outcomes. An answer can name your company without linking to you. It can cite your guide without recommending your product. A person can read either answer without visiting the site.
Use a separate record for each outcome:
| Outcome | What to record | What it helps you decide |
|---|---|---|
| Brand mentioned | Prompt, engine, market, date, whether your brand was named | Which buyer questions need attention |
| Your page cited | Exact source URL and the question it supported | Which content earns references |
| Qualified visit | Landing page, known referral source and a useful action | Whether discovery produces business value |
Choose one buyer group before building a prompt list. A useful starting panel might contain 20 questions about discovering a category, comparing options, checking a requirement and solving a practical problem. Twenty is a manageable planning choice, not a statistically sufficient sample for every business.
For a headphone brand such as JBL, these could include “What are the best noise cancelling headphones for a long-haul flight?” and “Which earbuds stay in while running?” Keep branded questions in a separate group: asking whether JBL headphones are good value is a different test from seeing whether JBL appears unprompted.
Write down the questions before you start editing pages. Our guide to AI visibility explains the distinction between visibility and traditional rankings.
Remove access problems before rewriting content
Check your homepage, your main product page and one useful guide. Each should return its real content without a login or a browser challenge. Inspect the first HTML response, the robots rules, the canonical URL and any noindex directive. Test the page's images too: a rule that blocks a short path prefix can accidentally block a longer public filename.
Official OpenAI documents OAI-SearchBot as its search crawler and GPTBot as a separate training control. Perplexity recommends allowing PerplexityBot and its published IP ranges. Anthropic distinguishes Claude-SearchBot, Claude-User and ClaudeBot. Treat those purposes separately when choosing access rules; training permission is not a substitute for search access.
An allowed robots rule is only one layer. A firewall can still block or challenge the request. Review actual verified crawler traffic if your hosting provider makes it available. Changing your own request's user-agent is a useful diagnostic, but does not prove that the provider's real crawler can reach the site.
Read the AI crawler robots guide for the individual controls and the JavaScript guide for checking what is present before the browser runs scripts.
Give each important question a useful destination
Create a question-to-page map before commissioning new articles. Put each buyer question beside the best existing URL, the answer missing from it, and the evidence you can add. Several related questions can belong on one substantial page.
Official Google's guidance does not require special AI files or a prescribed writing format. Its AI features build on Search eligibility; access does not guarantee selection. Treat the following editing checklist as a reader-focused workflow, not a ranking formula.
For each priority page:
- State what the product or topic is, who it serves and the specific problem it addresses.
- Answer the page's main question early enough that a reader can find it easily.
- Explain conditions and exceptions beside the answer, so an excerpt is less likely to mislead.
- Show evidence you can stand behind: an actual demonstration, a documented method, a reproducible example or an original measurement with its limitations.
- Link to the next useful step, such as a relevant tool, method or product detail.
On a headphone product page, “Long battery life” is too vague to help a buyer compare pairs. The hours with noise cancelling on and off, how long a short charge lasts and what the case adds are more useful. Do not invent a benchmark or a customer quote to make the page appear authoritative.
Keep commercial and informational intent distinct. A product page should explain what someone can buy and when. A troubleshooting article should solve the problem even if the reader never becomes a customer. Our query fan-out article explains how related questions can guide the scope of a single page.
Make your identity and evidence easy to verify
Keep the brand name, product description and availability consistent across the homepage, pricing page, documentation and public profiles you actually control. Give readers an about page and a clear publisher or author. Show when an article was published and update its date only after a substantive change.
Structured data should describe what a visitor can see. An Article or BlogPosting can identify its author, dates and publisher; the author can be an organization when that accurately describes the byline. This is a description of the content, not a certification of its quality.
For outside coverage, start with relevance. Find a small set of publications, communities and resource pages your buyers already use. Offer a reproducible finding, a useful free tool or an informed contribution that improves that resource. Record who covered the work, which page they linked to and what the coverage actually said.
This is an outreach hypothesis to test, not proof that any particular mention will affect an AI answer. Avoid bought recommendations, invented reviews and instructions hidden in pages that tell an assistant to recommend you. A public record worth checking is more durable than an attempt to steer an individual response.
Build a measurement workflow with free tools
You can establish a useful baseline without buying an AI tracking subscription:
- Google Search Console: inspect indexing and use its Search Generative AI performance reports for impressions of your URLs in AI features. Google announced these reports in June 2026 and says the rollout reached all websites by August 31. They are not a log of every prompt or brand mention.
- Bing Webmaster Tools: review AI Performance for page citations and grounding-query patterns across supported experiences. Its aggregated data is a sample, so it is not directly comparable to a count of your own test answers.
- Your existing analytics: inspect recognized AI referrals and useful actions on the landing pages. GA4 now has an AI Assistant channel; Google's documentation says this channel excludes AI Overviews and AI Mode.
- A spreadsheet: retain your fixed question panel, the environment used, the observation time, the answer's sources and your change log. Record failed checks separately from successful answers that did not mention you.
Manual checks have limits: availability, personalization and engine versions can change. If you run repeated observations, report the sample size and scope. Ten positive answers out of twenty successful observations is 50% within that small sample, not a claim that half of all buyers see you.
For more about interpreting repeated observations, read why AI visibility needs confidence intervals and how RookAtlas measures. Do not mix new questions into the old baseline without marking the change.
Run the work in four weekly passes
This is a suggested workload for one small site. Adapt it to the actual findings rather than treating every task as mandatory.
| Period | Work | Evidence to keep |
|---|---|---|
| Days 1–7 | Inspect access and indexing, choose the buyer group, save the initial prompt panel | URLs tested, access failures, baseline observations |
| Days 8–14 | Improve two existing pages and publish one missing guide | Before-and-after content, sources and a dated change log |
| Days 15–21 | Make one useful asset easier to find and share it with relevant people | Internal links, actual coverage and qualified feedback |
| Days 22–30 | Repeat the same observations, check cited pages and compare qualified visits | Comparable results, sample sizes and remaining uncertainties |
Change one clear group of pages at a time where practical. If you redesign the site, change the prompt panel and launch a campaign in the same week, attribution becomes much harder.
At the end, classify the outcome: access fixed, page indexed, citation observed, qualified visit observed, or still inconclusive. Those are different milestones. A clean technical check is progress even when recommendations have not changed yet.
Know what the result cannot prove
A before-and-after increase can coincide with your work without being caused by it. Other sites publish, engine behavior changes, and repeated answers vary. Compare like with like, retain unsuccessful results and avoid reporting only the prompts that improved.
Do not set a universal word count, keyword density or FAQ quota. Do not promise a citation because a page has valid schema. Keep an llms.txt index if another system uses it, but Google explicitly says it does not improve Google Search visibility.
Your next action should follow the evidence. If pages cannot be fetched, fix access. If they can be fetched but do not answer the question, improve the content. If answers use other sources, inspect what those sources provide. If the sample is too small to tell, collect comparable observations before spending more on a guess.
Questions
How long does it take to improve AI visibility?
There is no guaranteed timeline. Thirty days is a useful work cycle for fixing access, improving pages and collecting observations; crawling, indexing and answer selection may take longer or produce no measurable change.
Can I measure AI visibility for free?
You can start with Search Console, Bing Webmaster Tools, existing analytics and a fixed spreadsheet of manual observations. Each measures a different part of visibility, so keep impressions, citations, mentions and visits separate.
Does llms.txt make ChatGPT or Google recommend my brand?
No file guarantees a recommendation. Google says llms.txt does not help or harm Search visibility, and OpenAI documents OAI-SearchBot access as a control for search inclusion rather than a recommendation guarantee.
Should I publish a page for every AI prompt?
Start with one useful destination per distinct intent. Cover related questions together when that helps the reader, and create another page only when it answers a substantially different need.
Sources
- OpenAI: overview of OpenAI crawlers developers.openai.com
- Perplexity: crawlers and WAF configuration docs.perplexity.ai
- Anthropic: web crawlers and site-owner controls support.claude.com
- Google: optimizing for generative AI features in Search developers.google.com
- Google: AI features and your website developers.google.com
- Google: Article structured data developers.google.com
- Google: Search Generative AI performance reports (June 2026, rollout update August 2026) developers.google.com
- Bing: AI Performance bing.com
- Google Analytics: default channel definitions support.google.com



