Help customers find your services.

What does the work actually include?

The scope starts with your website and the questions your customers need answered. A useful improvement should leave you with something you can inspect, not an unexplained AI score.

• A page-by-page check of public access, redirects, canonical addresses and discovery rules.
• Clear answer destinations for real services and customer decisions, with gaps called out.
• Consistent public business facts, visible content and appropriate structured data.
• Separate provider checks for search discovery, model training and user-requested retrieval.
• A measured baseline and a record of the changes, tests and results that can actually be observed.

My website is online. Why is it not showing up in search?

Being online is the starting point. Check whether the important pages can be crawled, are indexed and explain the service a customer is searching for. A homepage that appears for your business name may still leave a specific service poorly explained. The next step depends on the gap: repair an access problem, improve an existing page or clarify the route to it.

Google Search and Gemini: check the right control

AI Overviews and AI Mode depend on Search eligibility and the site’s Search generative AI setting in Search Console. Gemini’s documented training and grounding uses have a separate Google-Extended control. Treating those as one switch can hide the real issue.

Google’s Search AI participation control

Google-Extended’s documented scope

ChatGPT: discovery is different from training

OAI-SearchBot is the search-discovery control. GPTBot concerns training, while ChatGPT-User handles certain user-requested visits. The checks need to match the purpose: allowing training is not proof that ChatGPT will find, use or cite a page.

OpenAI’s crawler controls

Claude: search and direct retrieval both matter

Claude-SearchBot and Claude-User serve different discovery and retrieval roles; ClaudeBot is the training crawler. Anthropic documents its own robots behavior and source-IP list. A request carrying a Claude name still needs verification before it is reported as genuine crawler activity.

Anthropic’s crawler guidance

Perplexity: verify the source, not just the name

PerplexityBot supports search rather than foundation-model training. Perplexity-User can retrieve pages for a person’s request and generally ignores robots rules. Their published address lists are separate. Neither a bot name nor an allow rule is proof that an answer used the page.

Perplexity’s crawler and retrieval guidance

Bing and Copilot: keep the source current

Bing’s discovery and indexing systems also support grounded AI experiences. Canonical links, a current sitemap and change notifications help communicate which pages are current. Preview and archive directives need care because they can restrict how content appears in Copilot. An accepted submission is a receipt, not an indexing result.

Bing’s webmaster guidelines

Does Bing register every backlink?

No. Bing’s backlink report is a representative sample of detected links, not a complete inventory. A business identity reference, a canonical relationship and a clickable link from another page are different things. An empty detail list does not prove that no links exist.

• Compare the exact referring page and destination, not just the business name.
• Keep your own published relationships separate from links observed by Bing.
• Treat sitemap submissions and IndexNow notifications as discovery messages, not backlink registration or ranking credit.

Bing’s backlink reporting limits

Mistral: three separate crawler roles

MistralAI-Index builds the search index; MistralAI-User retrieves pages for user questions; MistralAI-Training concerns model training. Their controls are independent. Mistral publishes address lists for Index and User, so their identity can be checked without a model API key.

Mistral’s publisher crawler specification

Grok and DeepSeek: separate documented support from assumptions

Keep accurate answers available on their public, canonical pages. The official documentation reviewed for these products did not establish a publisher-specific citation file or a verified crawler identity contract. A model API’s citation fields describe returned answers; they are not tags to paste into a website. Publisher support must not be held up by credentials for optional API tests.

Grok web-search documentation

DeepSeek’s developer documentation

What makes a page useful in an AI answer?

A clear answer to a real question, enough context to interpret it correctly and evidence for the claims. A repair page should explain when repair makes sense, what needs testing and what a customer should provide. A comparison should make the tradeoffs understandable. The format follows the question.

• State the answer clearly, then explain the conditions and limitations.
• Support factual claims with relevant sources or genuine, approved work examples.
• Keep the same facts available to people and automated readers.
• Give the reader a useful next step; do not create a page for every keyword variation.

Does every AI model need the same integration?

No. Separate provider modules should handle the controls, API requests, response formats and measurement that each product actually supports. A shared registry can connect those modules while the public answers and approved business facts stay consistent. That is different from maintaining competing copies of the business for each model.

• Write in the language your customers use; these pages use English.
• Use the provider’s documented data format for an integration, rather than inventing a special writing language.
• Keep useful answers visible on their canonical pages. Google explicitly does not require special AI files or special schema.
• Record the exact model and interface used in a test. A developer API result is not automatically a result from its consumer chat product.

Google’s AI-search optimization guidance

How do we measure whether it helped?

Keep access, visibility and visits separate. A verified crawler request shows retrieval activity. An indexed page shows index inclusion. A captured AI answer can show a citation. Website analytics can show an attributable visit. None of those observations alone proves an inquiry or a sale.

• Compare the same pages and reporting windows, with releases noted.
• Record the exact AI product, question, date and cited page when sampling answers.
• Keep a completed answer without a citation separate from an unavailable test.
• Report the limits alongside the results. Rankings and AI source selection remain the provider’s decision.

Bring your starting point

Share the website, the services you want it to explain and the questions customers ask. Existing search reports can help if you have access to them. A lack of account data is a gap to acknowledge, not a reason to invent a baseline.