I recently led a webinar for AMEC. It seems every AI presentation seems to arrive with a slides showing statistic so large that nobody is expected to inspect it, but everyone obeys and takes a photo for their Linkedin. In my AMEC Fundamentals Plus deck, I offered 350bn prompts, a $4.2tn opportunity and 112% transformation, all by next Tuesday. 

The figures were invented. That was the point. 
 
I prefer to focus on actionable insights. 

Generative engine optimisation (GEO) has already gathered its own crop of huge claims and seductive dashboards. The pressure to show a brand’s visibility in AI-generated answers is real. But a visibility score cannot tell you whether communications worked. 

That is why I chaired a processes with AMEC, the International Association for the Measurement and Evaluation of Communication, to develop and published seven GEO principles and a companion practitioner’s guide in May 2026. They apply familiar communications discipline to the way organisations are found, interpreted, cited and represented in AI-led discovery. 

The principles are a measurement discipline, rather than checklist of AI-search tactics, and it is great to see their adoption by our industry and how they are shaping PR in agency and in house teams. Their central instruction is pretty clear, measure the whole information environment online (yes media coverage, social media and owned content). Assess three evidence domains together – the public record, search and content readiness, and downstream AI outputs – using upstream evidence to interpret what appears in answers. Then connect those findings with evidence of awareness, trust, behaviour and organisational outcomes. 

AI discovery is a measurement problem 

AMEC isn’t here to wade into the GEO maelstrom and prescribe another set of optimisation tactics. Its role is to bring objectives, stakeholder information needs, transparency and outcome-led evaluation to a field that could otherwise be defined by vendor dashboards. 

As of July 2026, Google is rolling out a generative AI performance report to a subset of website owners, showing impressions and linked pages for AI Overviews and AI Mode. Bing’s public-preview AI Performance dashboard reports displayed citations, cited pages and sampled grounding queries. Specialist platforms can sample prompts, products and cited sources and, depending on the platform, preserve the resulting answers. These are of course lovely and useful diagnostics, but they observe different events and none proves communications impact. 

A score can help identify a problem or track a defined sample. It is not the communications objective, and it does not prove impact. Before asking how visible a brand is, we should ask what stakeholders need to know, what decision we hope to influence and what evidence would show progress. 

The seven AMEC GEO principles 

The AMEC GEO principles set a common standard for responsible measurement. In brief, they ask practitioners to: 

• measure AI-led discovery against communications objectives and stakeholder information needs  

• assess the upstream information environment before interpreting AI-generated answers 

• examine whether reliable information can be found, understood and cited 

• treat observed AI outputs as directional evidence and test them transparently across products, prompts, markets, languages and time; 

• distinguish visibility from outcomes, then connect AI discovery with evidence of awareness, trust, behaviour and impact; 

• give reliable, trustworthy and current sources greater weight than volume, promotion or short-term visibility; and 

• improve the public information environment without manipulating reviews, disguising promotion or flooding the web with poor material. 

The principles work together. The official wording and minimum evidence standards are available in the AMEC GEO Principles, while a handy A Practitioner’s Guide to GEO Measurement explains how to plan a programme, select prompts and products, document methods and report findings without false precision. 

Visibility is not an outcome 

An AI answer can tell us whether a brand appeared, how it was described and which sources were shown to the user. It cannot, on its own, establish whether anybody became aware of the brand, trusted it, changed an opinion or took action. 

Between an AI answer and an organisational outcome is an evidence gap. Audience research, referral behaviour and outcome data are needed to show whether visibility contributed to awareness, trust or action. 

Fifty prompts are not a market 

There is no known universe of prompt data. A set of 50 prompts may be a useful diagnostic sample, but it is not a census of stakeholder behaviour. 

Answers can vary by product, model, location, language, account history, context and time. Even small additions to a prompt can change the response. Any audit should therefore state which prompts were used, why they were selected, which products and markets were tested, how often the tests were run and what the sample cannot represent. 

Search terms and prompts also express intent differently. A search for “best crisis PR agency” is compressed into four words. A prompt might explain that a UK corporate communications team is preparing for regulatory scrutiny and needs advisers who understand media, policy, search visibility and evaluation. The second query asks for judgement from an LLM within a situation, not a page of ranked links. Totally different outcomes really. 

Keyword-style testing can be a starting point. For corporate reputation, issues and trust, it is rarely enough. 

Nor should ChatGPT, Claude, Google AI Mode and the Gemini app be rolled into one universal score. They can show different source and citation patterns in recorded tests, and those patterns can change over time. Test the products your stakeholders are likely to use and record the date, model, market, language and method. 

There is no durable shortcut 

Technical SEO remains necessary. Reliable information must be crawlable, indexed, clearly structured and current if search and answer products are to retrieve it. But technical work cannot manufacture independent authority. 

Google’s current guidance is unusually direct. Standard SEO practice still applies to generative search. Google says there is no special AI schema, no need to divide every page into artificial chunks and no benefit in creating an llms.txt file for Google Search. It says pursuing inauthentic mentions is unlikely to help. Separately, scaled content created chiefly to manipulate results can breach its spam policies. 

There is no need to declare war on every listicle. The problem is manipulation at scale, low-value linking and material produced chiefly to game a system. Our recommendation is to invest in useful original evidence, expert commentary, authentic discussion and independent validation. 

“GEO is not SEO with a cowboy hat.” That’s what I said in one of the presentations I made since the launch, and while that was a joke it is true.  PR helps build the public record through earned coverage, expert commentary, reviews, public documents and stakeholder discussion. Search specialists make reliable information retrievable. Analytics and research teams test what people do and think next. No single discipline owns the whole answer. 

Measure the whole information environment 

A credible GEO programme triangulates AMEC’s three evidence domains – upstream reputation, search and content readiness, and downstream AI outputs – before testing their relationship with behaviour and organisational outcomes. 

The public record 

Begin by defining the universe: the brand, its relevant competitors, stakeholder groups, topics, markets, languages, time period and accessible sources. 

At PR Agency One, OneEval Brand analyses accessible evidence across earned, shared and owned sources, adding paid sources where they are relevant and observable. It applies consistent collection and classification rules, comparing the brand and agreed competitors by share of voice, message delivery, brand attributes, sentiment and framing, and source authority. 

This shows what information is available around the brand. It does not, by itself, measure stakeholder perception. Claims about awareness, understanding or trust still require surveys, interviews or other audience research. 

Search and content readiness 

Reliable evidence should be easy to find, understand and cite. That includes technical basics such as crawlability, indexation, page performance and a clear site structure. It also includes consistent facts, well-labelled research, accessible reports, useful executive commentary and credible third-party support. 

Structured data can help a search engine interpret a page and qualify it for existing search features. It does not create authority or guarantee that an AI answer will cite it. 

AI outputs 

Downstream testing should record presence, framing, accuracy, omissions, recommendations, citations, source quality and reputational risks. Prompt sets should reflect real stakeholder questions, and the method should be repeatable so results can be compared over time. The report should state exactly what was tested and what the sample cannot show. One observed answer is not proof of total AI visibility. 

Start by separating five events that are often bundled into a single visibility score: 

• Grounding – the system retrieves information to inform an answer; the user may never see the source. 

• Citation – the answer displays a source as attribution. 

• Mention – the brand is named in the answer. 

• Recommendation – the answer advises the user to consider or choose the brand. 

• Referral – the user follows a link to the brand’s website. 

None guarantees the next. A source can inform an answer without being cited. A cited page does not guarantee a brand mention. A mention may be neutral or critical. A recommendation may never produce a click. 

Citation is the bibliography, not the brainstorm. Each observable event needs its own measure, and unavailable grounding data should be declared before we examine relationships with reputation, behaviour or business results. 

Behaviour and outcomes 

Since May 2026, Google Analytics 4 has classified recognised AI-assistant referral clicks under a dedicated AI Assistant channel. That is useful, but GA4 measures the click, not the answer. 

Match AI referral sources and landing pages with URLs cited or linked in saved answer tests or specialist monitoring. Then follow engaged sessions, key events, enquiries or sales, connecting the data with the customer relationship management system where possible. Read that evidence beside AI mentions, recommendations, branded demand and audience research. 

A citation without a click never appears in GA4,. Referrer data can also be lost, and analytics does not reveal the prompt or answer that preceded a visit. The evidence may support an association or contribution; it does not automatically prove individual attribution or causation. 

Ten practical rules for communications teams 

The principles can be converted into ten rules that any communications team can use. 

• Define the stakeholder question. Do not begin with a visibility score. 

• Set up brand tracking. Do not interpret AI outputs without the public record. 

• Define the relevant attributes and compare like with like. Do not default to global giants that bear little resemblance to the organisation. 

• Choose prompts and products according to the audience. Do not treat AI search as one channel. 

• Establish a baseline early and log every run. Do not rely on ad hoc screenshots. 

• Separate grounding, citation, mention, recommendation and referral. Do not collapse them into one score. 

• Audit the public record. Do not rely on owned pages alone. 

• Make reliable information findable. Do not chase schema theatre, link networks created to manufacture authority or scaled tricks. 

• Check accuracy, framing and recommendation. Do not count appearance alone. 

• Connect visibility with referrals, branded demand, audience research and organisational outcomes. Report contribution or association unless the research design supports a stronger claim. 

A useful direction is better than false certainty 

GEO measurement will improve as platforms disclose more data. It will not become credible through a larger score or a longer prompt list. 

Useful findings come with stated limits. Define the objective; examine the public record, search and content readiness, AI outputs and behaviour; document the method; explain the gaps; and test the relationship with reputation and organisational outcomes. 

The AMEC principles provide the discipline. Applying them begins with a simple refusal: do not report visibility as though it were impact.