AI visibility: where businesses should start

My starting point for AI visibility is the customer’s decision. What are they trying to buy, compare or understand? What would they need to know to choose your business? Where could an AI answer help or hinder that decision?

Answer those questions, then check whether the relevant services can access accurate, useful information about you. Establish what you can measure before investing in further changes.

That gives you a practical brief for your website, content and analytics teams. It also gives you something better to discuss than whether a chatbot mentioned you once.

This guide is for marketing and digital leaders. You can lead the work yourself, with help from whoever manages your website and reporting. The output is a short, prioritised action plan. Detailed platform configuration belongs with the person responsible for that platform.

You may hear this work called answer engine optimisation (AEO) or generative engine optimisation (GEO). Here, the practical job is to help relevant services find and accurately represent the business, then examine what happens next.

Start with a decision that matters

Choose one offer, one market and one customer decision. Use questions from enquiries, sales calls, customer research and search data. Include questions that don’t mention your business: a customer building a shortlist may not know your name yet.

Throughout this guide, imagine a hypothetical commercial cleaning business serving Sydney offices. These are illustrative questions, not findings from a real client:

Customer question	Information the business should be able to provide
Which cleaners service offices in Parramatta after hours?	Actual service area, office-cleaning offer and operating arrangements
What should an office-cleaning contract include?	Scope, exclusions, quality checks and how changes are handled
How do I compare cleaning quotes?	Factors affecting price, assumptions and useful questions to ask
Does this provider have experience with buildings like ours?	Relevant, permissioned examples and verifiable credentials

Pick the questions closest to a useful business outcome. For this example, that might be a qualified request for a site assessment. Write down what makes an enquiry qualified; otherwise, the reporting will reward any form submission that turns up.

Choose the AI services to investigate using customer evidence and available reporting. You can expand the scope later. A small, relevant starting point makes it easier to see what needs fixing.

Adapt the questions to your business. A retailer may need accurate variants, stock, delivery and returns across product pages and feeds. A B2B supplier may need to explain integration requirements, implementation and situations where its offer is unsuitable. Ask sales or operations which missing detail most often delays a decision.

Understand what visibility means

AI discovery can involve several operations. Crawling fetches a page. Indexing makes information available in a search system. Retrieval selects material to help answer a particular request. The answer may mention a business, cite a source or recommend a supplier. A person may then visit, enquire or buy.

The system generates that answer and may combine retrieved material with knowledge learned during training. Some answers use no live search. Updating a page changes an available source; it does not directly update the model or guarantee a corrected answer. Research on retrieval-augmented generation

These are separate observations. A crawler visit cannot establish that a customer saw you. A linked citation might support a general explanation without recommending your service.

Google describes its generative Search features as retrieving pages from its search index and using related queries to gather supporting material. Existing search foundations therefore remain relevant. Google’s generative AI guidance

The source material can also extend beyond your website. Keep a list of the business profiles, listings and other records customers encounter. Google identifies Business Profiles and Merchant Center as relevant sources for local and product information. Google’s business-information guidance

This guide focuses on web discovery and retrieved information. Model training is a separate use of content, with its own permission questions. As agents take on more evaluation and purchasing tasks, the same discipline—clear information, deliberate permissions and observable outcomes—provides a useful foundation for further work.

Decide what access you want

My view is that access should follow a commercial decision. What content or data are you prepared to give an AI service, for which uses, and what do you want in return? More access is only useful if the exchange makes sense for the business.

For a publisher, the content may be the product. If an AI answer substitutes for a paid article or a visit that funds the work, more extraction can undermine the business even when citations increase. I would weigh discovery against subscription value, advertising revenue and the possibility of paid access or licensing. A citation alone is not a business model.

A business seeking leads may reach a different conclusion. Making accurate service information available could help a prospective customer choose it. The return might be a qualified enquiry or a booking. That still leaves a decision about which material to expose and whether model training offers anything the business values. Public service details, paid research and customer records do not need the same policy.

This is the value-exchange argument I develop in The agentic layer: permission, attribution and a return to the business supplying the information should be deliberate choices.

Give your website owner a written policy naming the content, intended uses, expected return and person responsible for reviewing the decision. Consider search discovery, an agent fetching information for a customer and model training separately. Then check which distinctions the provider’s controls actually support.

The controls depend on the provider:

OpenAI: OAI-SearchBot supports ChatGPT search; GPTBot covers content that may be used for model training. Their permissions are independent. ChatGPT-User handles certain user-requested visits and does not determine search inclusion; OpenAI says robots.txt rules may not apply to those visits. OpenAI crawler documentation
Google: check the site’s Search generative AI inclusion setting in Search Console, including any inherited setting. Inclusion is the default. This covers specified Search and Discover features and is separate from training controls. Google’s documentation says the control reached all websites worldwide on 31 August 2026. Search generative AI control
Meta: Meta-WebIndexer supports Meta AI search. Meta-ExternalAgent covers uses including foundation-model training and indexing; it is not a training-only switch. Meta-ExternalAds serves advertising and business uses. Apply robots.txt preferences to the relevant crawler. Meta-ExternalFetcher handles user-requested fetching and agentic functions and may bypass those rules. Keep FacebookExternalHit, used for shared-link previews, separate; Meta also documents a security-check exception for it. Meta crawler documentation
Cloudflare: current controls distinguish Search, Agent and Training behaviour. A crawler can have more than one behaviour, so review which services a policy affects. Cloudflare’s bot classifications

Meta’s Commercial Terms preview, due to take effect on 30 October 2026, is another reason to examine the deal. The English (US) version checked on 20 October includes associated-website collection for advertising and business products, plus integrity and security. Review applicable terms alongside crawler settings. I discuss the wider issue in Meta wants your website in the bargain.

Treat robots.txt as published instructions to crawlers. It does not enforce access or secure private information. Google’s robots.txt documentation

For the cleaning business, a proposed policy might permit discovery of public service information while reserving training permission. The business owner must decide whether that policy fits the actual services and controls involved. It should be recorded before someone changes a setting.

Check what systems can actually access

Ask the website owner to inspect a few important pages: the main service page, service-area information and a page containing supporting evidence. Give them the URLs and the intended permission policy.

The task is to check whether the relevant provider can obtain the useful content, and whether the page meets that provider’s inclusion requirements. A page loading in your own browser is only one part of that check.

Ask for findings on:

Crawler instructions and indexing or snippet restrictions in page markup and response headers.
The response and content returned, including any login, error or security challenge.
Whether important information appears in the content the service can process, including pages that rely on JavaScript.
Available indexing evidence, kept separate from current access tests and verified crawler visits. Record the service, URL, method, date and remaining gaps.

For Google’s generative Search features, pages must be indexed, eligible for a search snippet and included through the relevant Search Console control. Eligibility does not guarantee appearance. Google’s technical requirements for these features

For Google, ask for the URL Inspection indexed result and a separate live test, including the content and selected canonical—the URL Google treats as the main version. A live-test pass does not establish indexing or ordinary Googlebot access. URL Inspection, Google’s crawler roles.

For Cloudflare users: review AI traffic settings alongside the web application firewall, which filters incoming requests. Cloudflare documents that a crawler allowed in AI Crawl Control can still be blocked by an earlier firewall rule. Ask for evidence from the affected request and rule before changing anything. AI Crawl Control and the WAF

Keep legitimate security protections in place while resolving a specific conflict. Ask the specialist to document the change, verify the result and retain a way to reverse it.

Use pass, needs attention, unknown or not applicable for each check. A pass needs evidence for the named service and page. If genuine crawler activity cannot be verified, mark that check unknown. Keep separate results for permissions, current content and provider-reported indexing. A request using a crawler’s name alone does not verify the provider’s identity; OpenAI, for example, publishes crawler IP ranges for this purpose. OpenAI crawler documentation

Make the business easier to understand and substantiate

Review the same pages against the customer questions you selected. My recommended starting checks are straightforward:

Explain the offer. State what you do, who it is for, where it is available and any important exclusions.
Help someone evaluate it. Explain the buying process, relevant options and factors affecting price. Publish useful boundaries where a fixed price would be misleading.
Support the claims. Use evidence you can stand behind: named expertise, permissioned case studies, clear methods, real examples and appropriately sourced data.
Make the information easy to use. Use descriptive headings, readable text and tables where comparisons help. Keep important conditions beside the claims they qualify.
Check consistency. Compare the website with the business profiles and listings you control. Investigate meaningful contradictions in third-party sources and request corrections where appropriate.

Examine important claims in your answer sample against current business facts and the linked sources. Does the source support the claim, including its conditions? Save the claim, source and correction route. Unsupported and false are different findings; inaccessible evidence is unknown. Visible citations need not show every source consulted. Research on citation support

For our hypothetical cleaner, “tailored solutions and exceptional service” gives a prospective customer little to assess. A useful service page could explain the offices served, available hours, what a standard visit covers, how the quote is calculated and how quality is checked. Every detail must reflect the actual offer.

Start by improving the existing page that should answer the question. Add a separate guide when the customer needs a fuller explanation. This is a recommendation about usefulness and clarity; it is not a promise of more citations.

Link related service and evidence pages with descriptive, crawlable links. Keep any existing structured data consistent with the visible content. Google’s link guidance, structured-data policies.

Treat specialist optimisation claims with the same discipline. Google says it does not use llms.txt to improve Search visibility and requires no special schema for its generative features. Those statements concern Google; other consumers need their own evidence. Google’s guidance on proposed optimisation techniques

For any proposed technique, ask which system consumes it, what problem it solves and how you would evaluate the result. Ask whether its evidence came from changing a live website or supplying documents directly to a model, and whether it measured citations, visits or sales. Those are different tests.

Establish a baseline you can interpret

Save your baseline before making changes. Record the period, market, product, filters and metric definitions. Keep these forms of evidence separate:

Evidence	What it can tell you	Boundary
Access checks and verified logs	Whether specified requests obtained content	Access alone does not establish appearance
Platform reports	Activity within the provider’s reported coverage	Definitions and coverage differ
A recorded sample of answers	How the business appeared in those runs, including inaccuracies	The sample is not the whole market
Identifiable visits and business outcomes	Recorded visits, enquiries or purchases under your attribution setup	They do not capture every influence on a decision

Google Search Console now has a generative AI performance report covering AI Overviews and AI Mode impressions. Its documented breakdowns include pages, countries, devices and dates. Google says it rolled out worldwide on 31 August 2026; insufficient impressions or exclusion can explain a missing report. These are appearances of links to the site, not every brand mention, an enquiry count or an AI-specific click count. Google’s report definitions

Bing Webmaster Tools reports citations across Microsoft Copilot, Bing’s AI summaries and selected partner integrations. Its grounding queries are retrieval phrases, not necessarily the customer’s original words; the data represents a sample of citation activity. A citation is a displayed source reference, not a click or a recommendation. Bing’s AI Performance definitions

These reports are evolving. Bing announced a Topics Portfolio beta on 13 October 2026, with approximate aggregates and data thresholds. Review new features against the decision you need to make before adding more reporting. Bing’s Topics Portfolio announcement

In website analytics, inspect identifiable sources alongside meaningful outcomes. GA4’s Traffic acquisition report includes Session source / medium and key events. Have the analyst check what the configuration actually captures. A key event is not automatically a unique, qualified lead; reconcile the agreed event with enquiry quality. GA4 Traffic acquisition

Missing referral information can leave traffic classified as direct. You cannot simply reassign that traffic to AI. Keep an optional customer-reported discovery question as a separate source of evidence. GA4’s direct-traffic explanation

For an answer sample, save the exact questions, product and interface, displayed model if known, date, language, location, session context, answers and citations. Record automatic or forced search and whether it visibly ran; use unknown when that is not exposed. Keep API and customer-facing app tests separate unless you have demonstrated equivalence.

Define counting rules before collecting answers. Include repeated runs and retain the same core questions and repeat weighting for comparisons. Keep branded and unbranded questions separate. Count mentions, linked sources, supplier recommendations and factual errors separately. Show counts beside percentages, and distinguish a valid answer with no mention from an unanswered request or a technical failure. The worksheet’s sampling example shows how.

Our cleaner could track whether sampled answers describe its service area accurately and whether identifiable visits produce qualified site-assessment enquiries. Those are useful observations even when a complete account of AI’s influence is unavailable.

An increase after a page edit is a finding to investigate. It does not establish that the edit caused it: demand, platform changes and other marketing activity may also have changed.

Turn the findings into work

Use the AI visibility action plan to assign each finding an owner, a specific action, completion evidence and a review date.

Start with commercially important information that is wrong, missing or unintentionally inaccessible. Establish the measurement needed to judge subsequent work. Put speculative techniques into a small, defined experiment once the foundations are understood.

For a small business, I would begin with one service, a few useful customer questions and the pages that should answer them. Review the work after implementation; choose a later observation window that fits your traffic and sales cycle. There is no universal deadline by which an AI system must start citing you.

Assign someone to revisit platform guidance and consequential business changes. Review the questions too: a new service area or buying process can make yesterday’s accurate page unhelpful.

The first milestone is a completed set of checks, clearer information and a baseline the business can explain. Use that evidence to decide the next investment. The review may also conclude that the foundations are sound and further AI-specific spending is not yet justified.