GEO sentiment needs more than a score

A conversation about GEO measurement this week reached an awkward point.

We could discuss whether an organisation appeared in AI-generated answers, how often it appeared and which sources were being cited. But presence alone could not answer the more commercially interesting question: what story was the system telling about the organisation once it had found it?

A brand can be highly visible and still be described as expensive, difficult to implement, suited to a different audience or behind the alternatives on a feature that matters. It can also be portrayed positively for reasons the organisation does not recognise or cannot substantiate. In both cases, the visibility metric looks healthy while the meaning of that visibility remains unclear.

This is where the idea of GEO sentiment becomes useful. It is also where a simple positive, neutral or negative score stops being enough.

Visibility can carry the wrong story

Most early GEO measurement has understandably focused on presence. Does the organisation appear for the questions that matter? Is it included in a comparison? Does the answer cite one of its pages? Is its share of visibility improving against a set of competitors?

Those are sensible starting points because an organisation that is absent cannot be considered. The problem is that presence measures inclusion, not interpretation.

An AI answer may name a business as a leading option and then attach a caveat that changes the recommendation. It may place the organisation in the wrong category, repeat an old limitation, associate it with a customer segment it has moved beyond or treat a minor capability as its defining strength. The brand is present, but a partial or outdated version of it is doing the work.

Being cited is not the same as being chosen. It may not even mean being understood.

That creates an uncomfortable possibility for reporting. Mention rate can rise while the organisation becomes more strongly associated with a story it would not choose for itself. The number improves, but the underlying representation may become less accurate or less useful.

I would be careful with the word sentiment

Sentiment is a tempting label because it gives us a familiar measurement frame. An answer sounds favourable, unfavourable or somewhere in between, and the results can be plotted over time.

But an AI system does not hold an opinion in the human sense. Its answer is a generated response to a particular prompt, produced by a particular model at a particular moment from whatever information and retrieval process shaped that interaction. A sentiment score describes a pattern in the output; it does not reveal what customers collectively believe, and it should not be treated as a brand-health measure.

Context changes the result too. The same organisation may be portrayed as a sensible choice for a new freelancer, a poor fit for a complex international business and an unnecessarily expensive option for somebody who needs only a basic feature. Those answers are not necessarily inconsistent. They may reflect genuinely different needs.

A single blended score would flatten those distinctions. It could turn a useful warning into false certainty by making several audiences, tasks, markets and comparison sets look like one stable reputation.

The better word, for me, is portrayal. Sentiment asks whether the language is positive or negative. Portrayal asks which version of the organisation the answer has assembled, for whom, and on what basis.

Presence, portrayal and proof

I think GEO measurement becomes more useful when it separates three layers.

Presence asks whether the organisation appears in answers to questions that matter. It should capture more than a raw mention: is the brand the default suggestion, one option in a list, a specialist choice, a cheaper alternative or an example to avoid?

Portrayal examines the attributes and cautions attached to that presence. Which audiences is the organisation associated with? What does the answer say it is good at? Which weaknesses recur? What comparison does the system keep making, even when the organisation does not frame itself that way?

Proof asks what public evidence supports or challenges that portrayal. Are the sources current? Do they come from the organisation, customers, independent reviewers or publishers? Is the answer relying on one repeated claim, or do several forms of evidence point in the same direction?

The distinction matters because each layer leads to different work. A presence problem may require technical accessibility, clearer entities or stronger distribution. A portrayal problem may expose confused positioning or an outdated narrative. A proof problem may reveal that the organisation keeps making a claim without supplying enough specific evidence for anybody else to repeat it confidently.

Collapsing all three into visibility gives the team a number without a diagnosis.

AI answers assemble a provisional reputation

I have argued that when search becomes recommendation, evidence becomes the interface. Portrayal adds another layer to that argument. Before somebody reaches the website, an answer may already have assembled a working account of what the organisation is, who it suits and what its limitations are.

That account does not come from one approved brand statement. It may draw on product pages, help documentation, reviews, comparison sites, customer stories, press coverage, public discussions and material written for different audiences at different points in time. The result can be coherent even when the underlying evidence is not.

This is why the assembled reputation is provisional. Change the prompt, model, retrieval source, language or date and the answer may change. The organisation should not pretend it can control that environment, and one report should not be treated as an objective reading of how AI sees the brand.

Provisional does not mean irrelevant, however. If people use these systems during research and evaluation, the portrayal becomes one of the contexts in which the organisation is encountered. The editorial question is whether that portrayal is specific, current and supported well enough to be useful.

That moves GEO beyond optimising individual pages. The unit of work becomes the evidence environment around the organisation, including the sources it does not own.

Not every portrayal problem is a content problem

This is the most important guardrail.

If an AI answer repeatedly describes a product as difficult to implement, the organisation should investigate the evidence behind that description. Perhaps old documentation is still prominent and clearer guidance would help. Perhaps several reviews describe a version of the onboarding process that has since changed, creating a legitimate case for current customer evidence.

But perhaps implementation really is difficult. In that case, publishing another article about simplicity would not solve the problem. It could make the public evidence more contradictory while leaving the customer experience untouched.

The same applies to unclear pricing, weak support, an awkward product limitation or a gap between the audience the company wants and the audience it actually serves. Sometimes the right GEO recommendation is not another article. It is a better product experience, a clearer price or evidence the organisation does not yet possess.

That is one reason GEO sentiment should not sit inside content reporting as a reputation-polishing exercise. A useful analysis routes the problem to the right owner. Some gaps belong with editorial or technical teams, while others belong with product, customer experience, pricing, brand, communications or senior leadership.

The observation only becomes intelligence when it changes the right decision.

Turn the score into a feedback loop

A practical approach would begin with a stable set of questions grouped by audience and need. The aim is not to find one prompt that produces an alarming screenshot, but to observe whether a portrayal recurs across relevant situations.

For each answer, the team should preserve enough context to interpret it: the prompt, model, date, market or language where relevant, the role assigned to the organisation, the claims made and the sources visible in the response. Without that context, a sentiment trend can look precise while mixing together unlike interactions.

The next step is editorial rather than statistical. Repeated portrayals can be tested against reality and separated into four broad states:

• accurate and useful, where the available evidence should be maintained and strengthened
• accurate but uncomfortable, where the issue needs to be addressed by the organisation rather than written away
• outdated, where current evidence exists but is difficult to find or has not travelled
• unsupported or misleading, where the answer needs clearer, attributable information that other sources can examine

That classification will never be perfect, but it creates a route from observation to action. It also makes the limits of the measurement visible. If an answer changes after an intervention, the team still needs to distinguish a durable shift from ordinary model variation.

This is where the GEO journalist becomes useful. The job is to inspect the answer, trace what may be shaping it, find the people who know whether the portrayal is fair and decide whether the next move is to renew, route, report or leave the evidence alone.

The aim cannot be positive answers at any cost

There is an obvious danger in this work. Once organisations can measure how favourably they are portrayed, they may start treating every critical answer as a problem to suppress and every positive phrase as something to manufacture at scale.

That would turn GEO sentiment into reputation laundering. It would also produce worse information.

A negative portrayal may be based on outdated evidence, but it may also be an accurate description of a trade-off. A product can be more expensive because it offers capabilities a smaller customer does not need. A service can require more implementation because it is designed for a more complex environment. Removing the caveat would not make the answer more useful; explaining the context would.

The objective should therefore be accuracy, not positivity. Good evidence can support a favourable claim, qualify an overstatement or confirm that the organisation has work to do. Editorial teams should be prepared for all three outcomes.

That is also why GEO should start with what the organisation knows. If the business cannot provide current product truth, customer experience and credible external proof, the content team cannot invent a better reputation without making the information environment less trustworthy.

A more useful GEO question

Visibility remains necessary. Technical GEO remains necessary too. An organisation needs to be discoverable, its entities need to be clear and its strongest evidence needs to be accessible to both people and machines.

But visibility is only the first layer. Once the organisation appears, the harder question is what the answer has made that appearance mean.

GEO sentiment can help expose that question, provided the score is treated as a signal rather than a verdict. Its value is not that it tells the organisation whether AI likes the brand. Its value is that it can reveal which version of the organisation is travelling, where that version came from and whether the evidence justifies it.

That is a narrower claim than saying a company can manage its reputation inside AI systems. It is also a more useful one. The task is not to force every answer to sound positive, but to make the organisation easier to represent accurately — and to notice when the answer reveals a problem that content alone cannot fix.