When search becomes recommendation, evidence becomes the interface

I saw a useful observation on LinkedIn this morning.

For years, when we wanted a recommendation, we asked somebody we trusted. Who is a good builder? Where should we eat? Which hotel should we book? Which company can actually help with this problem?

Search changed how we found the options, but there was still usually a gap between discovery and judgement. Google gave us pages, reviews, maps, directories and comparisons. We did the work of moving between them and deciding what looked credible.

AI search starts to change that interaction.

People can now ask the system a much more human question: which one should I choose?

That sounds like a small shift in phrasing. I think it creates a much bigger shift in what content has to do.

When search starts behaving more like recommendation, visibility is not enough. The harder question becomes whether there is enough evidence for an organisation to be a reasonable answer.

Search and recommendation are different jobs

Traditional search mostly helped with discovery.

That is a simplification, obviously. Search engines have been answering questions directly, ranking local businesses, surfacing reviews and shaping choice for years. Recommendation was already built into the system.

But the dominant interaction still left a lot of evaluation with the user. You searched, opened results, compared pages, checked reviews, asked somebody else and eventually made a decision.

A conversational system can compress more of those stages into one response.

“Find payroll software for a small business” is a discovery request.

“What payroll software would suit a ten-person UK company that wants something simple and already uses this accounting system?” is much closer to a judgement request.

The second question asks the system to interpret constraints, compare options and explain why one might fit better than another. That matters because the commercial challenge changes.

The question is no longer only whether the system can find us. It becomes what the system has available that would justify mentioning us in this particular situation.

I would not describe this as AI “trust”

This is where some of the language around GEO starts to bother me.

People say brands need to become businesses that AI “trusts” or “feels confident recommending”. I understand the shorthand, but I think it encourages the wrong mental model.

A model does not trust a company in the human sense. It does not have a relationship with the brand, remember a good experience or decide that the chief executive seems honest.

What it can have is information.

It can encounter claims about the organisation. It can retrieve documentation. It may have access to reviews, comparisons, customer stories, independent reporting, product information, public discussions and other sources. It can then synthesise those signals into an answer, sometimes well and sometimes badly.

That distinction matters.

If you tell a marketing team the goal is to make AI trust the company, it is easy to end up with another optimisation industry based on trying to send the right signals to the machine.

If you ask what evidence would make the company a reasonable recommendation for a specific need, the work becomes much more concrete. The questions become:

• What can we actually prove?
• What is specific enough to distinguish us?
• What do customers say happened?
• What do practitioners observe?
• What do independent sources confirm?
• Where does our own claim run ahead of the available evidence?

That is an editorial problem before it becomes a technical optimisation problem.

Recommendability is a useful way to think about it

I have been using the word recommendability in my notes because it captures something visibility does not. Visibility asks whether an organisation can be found and understood; recommendability asks whether the available evidence gives somebody, or a system helping somebody, enough reason to consider it a suitable answer in a particular context.

That is deliberately narrower than “being the best”. There is no universal recommendation.

A restaurant can be highly recommendable for a quick lunch near an office and completely wrong for a long anniversary dinner. Accounting software can make sense for one type of business and be a poor fit for another. A consultant can have strong evidence around one problem without being the right answer to an adjacent one.

This is one reason generic brand claims are so weak. “Trusted by businesses”, “industry-leading”, “built for growth” and “easy to use” may all be true, but on their own they give a recommendation system very little to work with. They do not define the situation in which the claim matters, the customer who experienced it or the evidence that supports it.

Recommendability needs context.

This is where the evidence layer becomes commercially important

I have already argued that content teams need an evidence layer, because AI makes competent production cheaper while first-hand knowledge remains scarce.

AI recommendation gives that argument another consequence.

Evidence is not only what makes an article better. It may increasingly become part of how a business is represented when somebody asks a system to compare choices.

I find it useful to separate several kinds of evidence:

• proprietary quantitative evidence, such as research or product data the organisation can genuinely stand behind
• practitioner evidence from people dealing with the problem in real conditions
• customer evidence showing what happened in a specific implementation or decision
• expert evidence that explains why something works, fails or matters
• external evidence from independent sources that can corroborate or challenge what the organisation says about itself

None of those categories is automatically superior to the others. A customer quote is not proof of a market-wide pattern. A survey can be statistically sound while telling you little about implementation. An expert can explain a mechanism without knowing whether customers experience it that way.

The value comes from knowing what each piece of evidence can support and where several different forms of evidence point in the same direction.

That is a much stronger foundation for recommendation than repeating the same owned claim across another ten pages.

The information environment matters more than the landing page

This is where the page-centric habits of SEO become limiting.

If somebody searches for a product category, the company understandably wants its relevant page to rank. The unit of optimisation is visible: a URL.

Recommendation is messier.

The answer may depend on information distributed across product pages, reviews, comparison sites, customer stories, technical documentation, press coverage, communities, public datasets and things other people say about the organisation.

The organisation does not control all of that environment, which is exactly why the evidence question matters. You can optimise your own statement that your product is easy to use, but you cannot directly manufacture independent agreement that it is easy to use.

You can publish a customer story saying implementation was fast, but you cannot assume one story establishes that as a general characteristic. You can create a beautifully structured page describing why you are suitable for small businesses, but you still need enough specific information for that suitability to mean something.

That is why I think AI search is optimising the wrong layer when it focuses too heavily on the page before asking whether the underlying information environment is strong enough.

The page still matters. It just is not the whole interface anymore.

Customer stories become more interesting in this model

This also changes how I think about customer stories.

A traditional case study has an obvious marketing job. It demonstrates that a customer used the product, experienced a problem, adopted something and got a result.

But a strong customer story is also a bundle of evidence.

It can show the starting condition, the constraints, why one option was chosen, what changed in the workflow, what did not work immediately, what improved and how the customer describes the result in their own language.

That is why I have been thinking about turning one customer story into an editorial system. The published page is only one use of the underlying material.

If the evidence is preserved properly, it can strengthen future answers to much more specific questions. Not “Why are we great?” but questions such as:

• What happens when a finance team outgrows this process?
• How long did implementation actually take?
• What did the customer have to change internally?
• Which problem was the product good at solving, and which problem still needed human judgement?

Those are recommendation-shaped questions. The closer content gets to answering them honestly, the more useful it becomes to humans as well as machines.

GEO should investigate the recommendation, not chase it

I think this suggests a healthier GEO workflow.

Start with the questions people are asking AI systems. Pay attention to the prompts that imply comparison, selection or judgement rather than only information retrieval.

Then look at the current answers.

Where is the organisation appearing? Where is it absent? What claims are associated with it? Which sources seem to support those claims? Where does a competitor have evidence that you do not? Where is the model making a distinction that your own content barely acknowledges?

The next move should not automatically be “create a page for that prompt”.

Sometimes that will be the right move. Often the more important intervention will be upstream. You may need:

• a better customer example
• clearer product constraints rather than pretending the product fits everyone
• independent research
• practitioner evidence about what actually happens
• an admission that the organisation does not yet deserve the answer it wants the model to give

That last possibility is uncomfortable, but it is also useful. GEO can become a way of identifying evidence gaps rather than only visibility gaps.

There is a strong reason not to overstate this

AI recommendation is still unstable.

Different systems can produce different answers. Models can be wrong. Retrieval varies. Commercial relationships may affect interfaces. Sources change. A recommendation can look authoritative while resting on weak or outdated information.

That makes “becoming the brand AI recommends” a dangerous strategic objective if it is treated literally, because it is too dependent on systems the organisation does not control. The more durable objective is to improve the information underneath the recommendation.

Make the product easier to understand. Publish specific evidence. Preserve customer proof. Be clear about who the offer is and is not for. Build external credibility where it can be earned. Correct weak claims. Make important information accessible and technically retrievable.

Those things remain valuable even if today’s answer engine disappears.

This is also why technical GEO still matters. If strong evidence exists but machines cannot retrieve or interpret it, that is a real problem. Structure, accessibility, entity clarity and monitoring all have a job.

The mistake is assuming they can substitute for the evidence itself.

The question has moved upstream

For years, the content question was often: how do we rank for this? AI search added another one: how do we appear in the answer? Recommendation adds a harder question: why would we be a reasonable answer in the first place?

That is the question I think content, communications, customer marketing, product marketing and GEO increasingly meet around. It pushes the work upstream from optimisation into research, customer evidence, product truth, external reputation and editorial judgement.

It also changes what I think the interface really is. In traditional search, the webpage was often the place where the organisation made its case. In AI recommendation, the case may already be partly formed before the person arrives.

The interface is not only the page. It is the evidence environment around the organisation, and whether that evidence gives a human or a machine enough reason to say: for this particular problem, this is worth considering.