Optimising for Gemini, Claude and Copilot

Three assistants, three routes to your site

All three assistants write answers from a mix of what their model learned in training and what they fetch from the web at the time of the question. The part you can influence quickly is the fetching, and that is where they differ.

The overlap is larger than the table suggests. All three reward the same thing once they have your page: a clear, specific answer near the top, written so a sentence or two can be lifted out and still make sense. The differences are mostly about getting found in the first place.

Gemini: competing inside Google Search

Google's own guidance on AI features says there are no extra technical requirements for appearing in AI Overviews or AI Mode beyond being indexed and eligible to show a snippet, and that normal Search best practice applies. That is useful to know, because it rules out a lot of paid advice about special AI tags. What it leaves is the quality and coverage of your content.

Fan-out questions

Google has described AI Mode as breaking a question into several related searches, running them together and combining what it finds. A person asking "Is it worth fitting a heat pump in a 1930s semi?" may trigger searches about running costs, insulation, grants, noise and installation time. A page that answers only the headline question competes for one of those searches. A page, or a small cluster of linked pages, that answers the follow-ups too has more chances to be used.

You can find likely follow-ups in the "People also ask" boxes for your main question, in Search Console queries that already bring impressions to the page, and in the questions customers ask you by phone and email.

Pages that already rank

For most sites the fastest gains in Gemini come from pages already on the first or second page of Google for the question. They are indexed and trusted; they often just bury the answer. Rewriting the opening of those pages so the answer comes first is usually worth more than publishing new ones.

Local questions

For "near me" and service-area questions, Google's answers draw heavily on business listings. A complete and accurate Google Business Profile, with the right categories, services and hours, matters as much as the website. See the Google Business Profile fixes for that side.

Copilot: the Bing side

Copilot builds its web answers from Bing. If Bing has not indexed a page, Copilot cannot cite it, and many sites that have spent years on Google have never looked at Bing at all. The work is quick and mostly one-off.

• Bing Webmaster Tools. Verify the site, and if you already use Google Search Console you can import the verified sites from it rather than verifying again. Submit your XML sitemap and check the indexing reports for pages Bing has excluded.
• IndexNow. Bing supports IndexNow, a simple protocol that lets a site tell search engines a page has been added, changed or removed. Many CMS plugins and CDNs can send IndexNow notices automatically, which shortens the wait for Bing to notice an update.
• URL Inspection in Bing. Bing Webmaster Tools has its own URL inspection, which shows whether a page is indexed and whether Bing found problems with it. Use it on the pages you most want Copilot to cite.
• Robots meta tags. When Microsoft launched chat answers in Bing in 2023, it said the noarchive and nocache robots meta tags would limit how page content is used in those answers. Search your templates for either tag; many sites added noarchive long ago for unrelated reasons. Check Bing's current webmaster guidance before relying on the exact behaviour.

<!-- Remove these from pages you want Copilot to use, unless they are deliberate -->
<meta name="robots" content="noarchive">
<meta name="bingbot" content="nocache">

Claude: Anthropic's crawlers and fetches

Anthropic documents three crawlers, each with a separate job, and lets site owners control each one by name in robots.txt:

• ClaudeBot collects content that may be used to train future models.
• Claude-SearchBot crawls pages to improve the results Claude uses when it searches the web.
• Claude-User fetches a page when a person using Claude asks about it or when Claude needs it to answer them.

Blocking ClaudeBot is a training decision and does not stop the other two. Blocking Claude-SearchBot or Claude-User reduces how often Claude can cite or read your pages when people ask questions today. A wildcard User-agent: * with Disallow: /, sometimes left over from a staging site, blocks all three.

# robots.txt: let Claude search and fetch; training decided separately
User-agent: Claude-SearchBot
Allow: /

User-agent: Claude-User
Allow: /

User-agent: ClaudeBot
Disallow: /

Claude often fetches a page directly and reads it as text, so pages that depend on scripts to show their content, or that put a consent wall or pop-up in front of it, can come back nearly empty. The test in Step 6 shows you what Claude actually gets.

The fix, step by step

Work that helps all three at once

Most of the effort in AEO is not assistant-specific. These tasks improve your chances everywhere and are worth doing before any fine-tuning for a single assistant:

• One page per question cluster. Several thin pages on the same subject split your signals. A single thorough page with clear sections is easier for any assistant to choose.
• Facts stated plainly. Prices, areas served, opening hours, specifications and timescales written in sentences, not only in images or downloadable PDFs.
• Named authors and a real About page. Who wrote the page, what qualifies them and how to contact the business help every assistant judge trust and attribute information correctly.
• Consistent identity. The same business name, description and contact details on your site and every profile, supported by Organization markup with sameAs links.
• Mentions elsewhere. Reviews, directories, trade bodies and local press are read by all three. See fixing weak brand mentions.

For writing technique in more depth, including before-and-after examples, see fixing content structure for AI extraction.

What not to spend time on

• Special AI meta tags. None of the three assistants documents a tag that increases your chances of being cited. Tags you find recommended online either do nothing or are standard robots rules that can only restrict use.
• Allowing training crawlers to "get cited". Training crawlers such as ClaudeBot and the Google-Extended token affect future models, not the live search that produces most links. Decide on training separately and on its merits.
• Rewriting for each assistant. Separate versions of a page for Gemini, Claude and Copilot are not possible and not needed. One clear page serves all three.
• Chasing a single answer. An answer that includes you today may not tomorrow. Judge progress over a month of repeated prompts.
• Keyword repetition. Repeating a phrase does not make a passage more quotable. Specific, correct facts do.

Measuring progress for each assistant

Traffic from these assistants is hard to see in analytics. Gemini's contribution through AI Overviews and AI Mode is counted inside the ordinary Web totals in Google Search Console. Copilot and Claude may pass a referrer when someone clicks a source link, but not always, so some visits show up as direct traffic.

That makes prompt-based measurement the main yardstick. The AI Visibility Tracker asks AI assistants the questions your customers ask and records whether your brand is cited, mentioned, recommended or missing, and tracks that over time. Keep the prompt set fixed so month-on-month comparisons mean something, and add new prompts as a separate group rather than mixing them in.