AI May Introduce Your Business Before Your Website Does Fix the Public Story First

A buyer may meet a generated description of your business before they meet your website. They may also meet a review, directory profile, search result, supplier page or competitor comparison first.

You cannot control which route appears or guarantee what every answer says. You can control whether the public facts available along those routes are current, compatible and supported.

My position is that this is a brand-governance problem before it is an AI-optimisation problem. The first website visit is no longer necessarily the first explanation.

The buyer can arrive with the wrong company already in mind

A prospect may reach the site already believing that the company serves a different market, provides a discontinued service or lacks a capability hidden behind vague language. Analytics records the eventual visit. It does not reveal the summary, profile or comparison that framed the buyer’s questions beforehand.

This creates what I call pre-visit positioning debt: the website has to spend the opening moments correcting an explanation formed elsewhere before it can make the actual case.

Sometimes the external description is wrong. Sometimes it is a painfully accurate summary of contradictions the business has left in public. One page calls the company a specialist consultancy. Another presents a general agency. An old profile advertises a service the team stopped delivering. A location page promises coverage the operation cannot support.

I would not start by blaming the answer system. I would ask whether an unfamiliar person could reconstruct the correct business from the public record without receiving the positioning deck first.

Your website is a source, not a model control panel

Claims that a website can “train AI to recommend the brand” collapse crawling, indexing, retrieval, generation, citation and buyer exposure into one convenient sales lever. A page may be crawled, indexed, retrieved, linked, summarised, quoted or ignored. Products can use different systems and sources. Editing a homepage does not give the business direct access to a model’s beliefs or guarantee a future mention.

Google says its normal SEO foundations remain relevant for AI Overviews and AI Mode. It states that no additional requirements, special AI files or special schema are necessary, while eligibility does not guarantee appearance (Google’s guidance on AI features and websites). OpenAI separately documents web crawlers and controls for different purposes, including search surfacing and potential model training (OpenAI crawler documentation). Access settings can permit or restrict a route. They do not configure a recommendation.

My evidence ceiling is therefore narrow: inspect public sources, technical access and dated answer samples. Do not turn one screenshot into proof of stable “AI understanding”, market exposure or commercial influence.

Write the eight statements the public story must preserve

Before auditing URLs, write the material story the business is prepared to stand behind:

• Identity which legal or trading organisation is being described.
• Category the ordinary market category that best explains the business.
• Audience who the offer is designed to serve.
• Offer what the business currently provides.
• Method the operating approach a buyer needs to understand.
• Proof what supports the consequential claims.
• Fit and limits where, when and for whom the offer is or is not suitable.
• Next action what a qualified buyer should do next.

These are not eight sentences to paste across every page. They are story invariants: facts and propositions whose meaning should survive changes in format, length and audience. The homepage, service page and business profile can use different words. They should not create incompatible identities, offers or limits.

For each invariant, record its maintained public home, evidence, accountable owner and change trigger. Otherwise a rebrand updates the presentation deck while old pages continue introducing the previous business.

Distinguish useful variation from a contradiction

Exact wording consistency is not the goal. Classify the difference:

I prioritise differences that can change a buyer’s decision: category, service scope, geography, availability, price conditions, qualifications, ownership relationships and meaningful exclusions. A homepage calling the company a specialist while the About page describes a general agency is not just tonal variation. It creates two categories. A page promising nationwide delivery while the operating team covers one region is not an SEO detail. It can generate the wrong enquiry before any AI system becomes involved.

Correct the fact where it lives

When the public story conflicts, repair the source closest to the fact:

• Correct the offer on the service or product page that owns it
• Put eligibility and exclusions beside the promise they limit
• Update address, service area, hours and contact data in the maintained profile
• Attach a qualification to the person or organisation that holds it
• Replace obsolete team, partner or ownership statements at their source
• Request third-party factual corrections through the publisher's process
• Leave independent opinion alone when it is not a factual error
• Retain unresolved facts as unknown until evidence settles them

Google’s Business Profile guidance asks businesses to represent themselves as they are consistently recognised in the real world and keep key information accurate (Google Business Profile guidelines). That is worth doing whether a generated answer has repeated the mistake or not.

I would not publish six new articles to overpower one stale profile. Volume does not repair source ownership. It gives the contradiction more places to spread.

Make every important entry page pass the Stranger Test

A visitor may first enter through a service page, article, case study or location page. That page need not reproduce the homepage. It must provide enough local context for an unfamiliar person to identify the source and understand the claim. For each important entry page, ask:

• Which organisation or person is speaking?
• What subject or offer does the page own?
• Which audience and situation does it address?
• What consequential claim does it make?
• What visible evidence supports that claim?
• Which limitation changes its meaning?
• What should the right reader do next?

I call this the Unbriefed Stranger Test. Give the page to someone without the brand deck and ask them to repeat the story. Compare their account with the eight invariants. The gaps are often more revealing than a readability score. People inside the company can decode “solutions”, “growth” and “end-to-end support” because the missing context is sitting around the meeting table. Outside that room, the words describe almost nothing.

The test does not prove what an AI product will say. It shows whether the public explanation asks any outside reader to invent essential context.

Sample generated answers to find defects, not certainty

Once the public story is mapped, sample answer products for commercially meaningful questions. Treat each output as an observation under recorded conditions. Capture the product and mode, exact question, date, locale, material account context, complete visible answer, linked sources and the brand statement being checked.

Compare that statement with the approved invariants. Look for the wrong identity, category, audience, offer, location, limitation or relationship. Then inspect the visible sources before changing the website. One inaccurate answer does not prove a market-wide brand problem. One correct answer does not prove stable understanding or recommendation. Repetition under documented conditions may justify deeper investigation, but it remains sampled evidence.

My decision when the public record is accurate and one output is wrong may be to monitor. Rewriting a truthful page to chase an unstable answer can make the source worse for everyone else.

Keep source, answer and buyer evidence apart

“AI first impression” becomes misleading when three evidence streams are blended into one score:

A screenshot does not prove buyer exposure. A salesperson’s recollection does not prove the exact wording of an answer. An accurate website does not prove inclusion. The streams can inform one another. They cannot be laundered into “AI mindshare” because a report needs a percentage.

I would report decisions the business can own: material contradictions unresolved, stale facts without owners, entry pages that fail the Stranger Test, repeated inaccuracies under comparable samples, human explanation gaps and corrections awaiting release.

Reopen the story when reality changes

Review the public story when a service launches, closes or changes scope; the audience or market category changes; a location, delivery area or availability condition changes; a merger, acquisition, rebrand or legal-identity change occurs; a credential, named expert or proof point changes; prospects repeatedly misunderstand a material part of the offer; a significant third-party record changes; or a material generated inaccuracy recurs under comparable conditions.

A periodic check can catch slow drift. It cannot replace factual ownership. Monthly answer screenshots are theatre when nobody is responsible for correcting the About page.

Before asking what AI “thinks” about the company, make the public business reconstructable. Define the identity, category, audience, offer, method, proof, limits and next action. Give each invariant a maintained home. Resolve contradictions where they originate. Test priority pages without the positioning deck. Sample answers without pretending to see inside the system.

The goal is not to make every source recite identical copy. It is to make the correct business easier to recognise and harder to accidentally rewrite.