How to improve AI visibility of a website

The Actions tab in the extension looks at your measurement and your site’s public files, then shows up to three next steps. Each one is labeled with a grade:

• A — an AI engine or search engine’s own documentation, or a peer-reviewed paper, directly supports it.
• B — official documentation lists it as a general best practice.
• M — confirmed by our own citation benchmark (see Page comparison): across real measurements, pages that have it are cited over pages that lack it more often than chance.
• O — observed in your own measurement (for example, a competitor’s page was cited for a question you have no page for). Not a general claim, a specific finding.

We only show A, B, M, and O grade items in the panel. This page also covers items that are worth doing but cannot be detected from outside a site, and items that are commonly recommended but not confirmed by any engine’s own documentation.

What we check on your site

• AI crawler access. Whether your robots.txt allows the crawlers that OpenAI, Perplexity, and Anthropic each document by name (OAI-SearchBot, PerplexityBot, Claude-SearchBot), or blocks them, or blocks everything.
• Indexing signals. Whether your home page’s meta robots tag or X-Robots-Tag header carries noindex, nosnippet, or none, which Google’s own documentation on AI features covers directly.
• Body rendered in HTML. Whether your main content is present in the HTML response itself rather than requiring JavaScript to render, per Google’s guidance to follow JavaScript SEO best practices.
• Canonical tag. Whether your home page has a canonical tag and it points to itself rather than a different host.

What we learn from the measurement

• Questions where a competitor is cited and you have no page. We compare the questions used in your measurement against your sitemap. When a competitor’s page was cited for a question and none of your pages answers it, that is a specific, observed gap.
• Evidence density on cited pages. A 2024 paper on Generative Engine Optimization (Aggarwal et al., KDD 2024) found that adding citations, statistics, and quotations to a page measurably increased how often it was selected in AI-generated answers, while stuffing content with keywords made things worse. The paper’s abstract reports visibility gains of up to 40% from these techniques. The evidence is mixed: the same paper found that pages already cited first lost visibility from these edits, and a 2025 benchmark (C-SEO Bench, NeurIPS 2025 Datasets and Benchmarks) found significant ranking gains in only 3 of 54 settings. So we flag it only when you are not the page being cited and a competitor’s cited page carries meaningfully more numbers and outbound references than your own page.
• The exact third-party pages the engine cited instead of you. AI answers lean on a handful of places per category: a community thread, a review or directory site, a comparison article, a video, Wikipedia. We keep the URL of every cited page, read the public page, and check whether it names you. When it does not, we show the page and a step that names the place and says how to do it: start a thread that answers the question in the subreddit whose thread AI cites, or reply in that thread; answer the cited Quora or Stack Overflow question; claim your profile on the cited review site; ask the publisher of the cited article to include you; publish a YouTube video that answers the question; collect customer reviews on the site that answers questions about your brand; or, when answers in your category lean on Wikipedia and it has no article about you, submit one through Articles for Creation. If a page cannot be read (many communities refuse automated requests), we say that the mention could not be checked rather than claiming you are absent.
• Disclosed participation only. We never suggest posting on your behalf, undisclosed posting, or manufactured mentions, and we do not assist with them. Google’s AI optimization guide says seeking inauthentic “mentions” across the web “isn’t as helpful as it might seem,” and its spam policies cover attempts to manipulate generative AI responses; genuine participation in the places an engine already cites is a different thing.

Page comparison

For each question where a competitor’s page was cited, we find the page on your site that answers it and compare the two on the same criteria. Where the cited page has something yours lacks, that becomes a task naming your page, for example “Cite the sources behind the claims on your pricing guide”.

The criteria.

• Answers the question directly (B).
• Cites sources (A), quotes a named person or organization (A), carries specific figures (A), and links to its sources (A).
• First-hand experience or original data (B), and a named author (B).
• Answers at the start of the page or section, shows when it was updated, has sections that stand on their own, explains its method or publishes its data, and has a comparison table. These are all C-grade advice, turned into tasks only once our benchmark has confirmed them (grade M).
• Keyword stuffing, as a control. It is expected to count against a page, and if it ever does not, we treat that as a fault in the benchmark.

The benchmark. Every measurement adds comparisons: for one question, the page the engine cited and your page it did not cite. Until these comparisons agree with real citations often enough, the store version of the extension does not show these tasks.

Checking a task. Marking a task done reads your public files and pages again. When the gap is gone, the task is marked verified. When a checkable task is still found, it says so; a change can take a moment to go live. The cited pages are not re-read; they are not yours to change. A check sends nothing new from your browser: the domain and the anonymous identifier, as for every request.

Worth doing, but we cannot see it from outside

• First-hand expertise and a distinct point of view. Google’s AI optimization guidance recommends original reporting, testing, and clearly stated authorship over recycled summaries. We cannot measure authorship from outside your site.
• Google Search Console’s Generative AI report. Rolled out worldwide since August 31, 2026, it shows impressions for AI Overviews and AI Mode, though not query-level data. Only you can see this, from your own Search Console account.
• Bing Webmaster Tools’ AI Performance report. In public preview since February 2026, it includes “grounding queries” — the phrases that retrieved your cited content. It covers Microsoft Copilot and Bing AI, and does not cover ChatGPT, Perplexity, Google, or Claude.
• IndexNow. A real, documented protocol for notifying search engines when your content changes.
• Google Preferred Sources. Readers can pick your site as a preferred source; in AI Mode and AI Overviews your content can then carry a “preferred” badge for them. Domains and subdomains are eligible, subdirectories are not.
• Merchant Center and Business Profile data. Listed among Google’s fundamentals for AI features, relevant if you sell products or run a local business.
• Internal links, images, and video. Listed among Google’s general fundamentals that also apply to AI features.

Commonly recommended, not confirmed by any engine

These are frequently suggested in GEO advice, but we found no official documentation or peer-reviewed source confirming they affect AI citation. Treat them as reasonable general practice, not as a lever we can verify.

• Opening a section with a direct answer to the likely question.
• Publishing on a regular schedule and keeping a visible “last updated” date.
• Writing sections that stand on their own and could be quoted independently.
• Using comparison tables where relevant.
• Publishing raw data or a methodology page alongside your conclusions.
• Using semantic HTML and ARIA roles beyond ordinary accessibility practice.

What official sources say does not help

• Adding structured data specifically for AI features. Google’s own documentation states there is no special structured data needed for AI features. Keeping your existing structured data accurate is still a general fundamental, but adding new schema for this purpose will not move the needle.
• llms.txt. Google states these files “aren’t needed for Google Search” and that Google Search itself doesn’t use them.
• FAQ schema for rich results. Google stopped showing FAQ rich results on May 7, 2026, and removed their documentation in June 2026.
• Fixed section lengths. Google states there is no ideal page length.
• Manufactured mentions. Google’s AI optimization guide says seeking inauthentic “mentions” across the web “isn’t as helpful as it might seem,” and Google’s spam policies include attempting to manipulate generative AI responses in Google Search.
• Hidden text aimed at AI systems. This falls under Google’s spam policies on cloaking and hidden text.

How to verify

Verification works in two layers, and they move at different speeds.

Detection changes in minutes. Fix something the Actions tab flagged — allow a crawler, remove a noindex tag, add a canonical — and mark it done. Your public files are read again right away; if the fix worked, the task is marked verified.

Citation changes take weeks, and need more than one measurement. AI engines re-crawl on their own schedule, so a single before-and-after comparison is not reliable. We recommend measuring several times before a change, making one change at a time, and re-measuring roughly 7, 14, and 28 days after. Judge the result by whether your score keeps moving over several measurements, and whether your citation gap against a specific competitor has narrowed for the same question — not by a single score reading. Actions never change your score by themselves; only a change on your site, followed by a new measurement, can.

The console does this comparison for you. For a tracked site, every action marked done is compared, three weeks before against three weeks after, with the sites in its market that did not do it, so what the engine did to everyone that month is subtracted. When your own change stands out of that noise, the Today feed says the action was verified. Across all participating sites, the same comparison is pooled per action and published as the benchmark: how many sites it worked for, by how much, and how long each engine took to reflect it. Fewer than thirty sites and it is shown as a trend; fewer than five and it is not shown at all.

Sources

• OpenAI crawlers: https://platform.openai.com/docs/bots
• Perplexity crawlers: https://docs.perplexity.ai/docs/resources/perplexity-crawlers
• Anthropic crawlers: https://support.anthropic.com/en/articles/8896518
• Google Search: robots.txt and Google-Extended: https://developers.google.com/search/docs/crawling-indexing/robots/intro
• Google: AI features and your website: https://developers.google.com/search/docs/appearance/ai-features
• Google: AI optimization guide: https://developers.google.com/search/docs/fundamentals/ai-optimization-guide
• Google: Spam policies: https://developers.google.com/search/docs/essentials/spam-policies
• Google Search Central: Latest updates (llms.txt, FAQ rich results): https://developers.google.com/search/updates
• Google: Preferred Sources: https://developers.google.com/search/docs/appearance/preferred-sources
• Aggarwal et al., “GEO: Generative Engine Optimization,” KDD 2024: https://arxiv.org/abs/2311.09735
• C-SEO Bench, NeurIPS 2025 Datasets and Benchmarks: https://arxiv.org/abs/2506.11097
• Wikipedia: Articles for Creation: https://en.wikipedia.org/wiki/Wikipedia:Articles_for_creation
• Bing Webmaster Tools: AI Performance (public preview): https://blogs.bing.com/webmaster/February-2026/Introducing-AI-Performance-in-Bing-Webmaster-Tools-Public-Preview
• IndexNow documentation: https://www.indexnow.org/documentation