Semrush Spotlight: Your website can't prove everything you claim

An external-evidence audit for GEO: find out who else can substantiate what you say, what AI answers are drawing on, and what an honest editorial response would look like.

A software company says that its product is easy to implement. The website has a reassuring headline, a polished customer film and a product page explaining the setup. Nothing is necessarily wrong with any of that.

But a prospective customer has a different question: how difficult is it really to get started? They might look for a review, watch somebody use the product, ask a professional peer or compare it with the alternatives. An AI assistant attempting to answer may surface some of those same sources, alongside the company’s own material.

What happens if the customer and the company tell different stories?

That question stayed with me after Semrush Spotlight in London. The conference brought together several lines of evidence about AI discovery, and Kevin Indig’s session on what he called the trust layer gave me a practical way to think about the material beyond a company’s own site. It also connected to a longer-running question in my Editorial Intelligence work: what is the point of making an organisation’s knowledge more discoverable if the claims themselves are not adequately supported?

I have already written about how search becoming recommendation makes evidence more important. That piece examined why a brand needs specific, credible information if it wants to be considered for a particular need. The question here is different and more operational: where is the evidence outside the organisation, who controls it, and how do you find out whether it is helping or contradicting the story you tell?

An information competitor may not sell a competing product

Traditional competitor analysis gives marketing teams a fairly understandable list: the businesses offering similar products to similar customers. SEO widens that list because a publisher, comparison site or government page can outrank a vendor for an important query.

AI answers make the distinction particularly visible. A company may be mentioned in an answer while the sources used to construct that answer belong to somebody else. The business competing for consideration and the source competing to explain the problem are not necessarily the same organisation.

The new Semrush 2026 AI Visibility Index, announced at Spotlight on 13 October, analysed seven months of results across ChatGPT, Gemini, Google AI Mode and AI Overviews, using fixed monthly prompt sets built from Semrush’s own search data and an additional variable dataset for AI Overviews. The UK brands that appear everywhere are mostly retailers, and the index was launched at Semrush’s own conference, so I’d treat it as evidence about the shape of the problem rather than a benchmark for B2B. Its UK findings underline two cautions.

First, there is no consistent source environment across platforms or markets. Among ChatGPT’s 100 most-cited domains, news publishers accounted for an average 29.5% of citations in the UK, compared with 19.6% in the US. That is a share within a particular top-domain group, not a claim that almost a third of every UK answer cites a news publisher.

Second, being mentioned and being cited are different outcomes. Semrush found that, averaged across 22 categories and four platforms, fewer than half of the top 100 mentioned brands also appeared among the top 100 cited domains in each month and market measured. A brand may be well known enough to name without its own pages being central information sources. Semrush’s own example is stark: Tesco and Asda were both among the top 100 mentioned UK brands on all four platforms every month, yet in July Tesco was ChatGPT’s 25th most-cited UK domain and Asda its 606th.

Neither result proves that every B2B company needs a large digital PR programme. It says something more basic: an accurate diagnosis of AI visibility has to examine the sources that answer the buyer’s question, not just the company’s rank against its commercial competitors.

Consider a UK financial-software question about regulatory reporting. The product competitors may be other software companies. The information environment may also include official government guidance, accountants explaining their actual routines, professional bodies, customer reviews, business publications and demonstration videos. Those sources are doing different jobs. Some explain the law; some reveal the experience of implementation; some compare products. A good answer may need more than one.

What the trust layer helps us inspect

At Spotlight, Kevin Indig described a four-stage way of approaching this: the claim a brand makes, the evidence available on third-party sites, the portrayal assembled in an AI answer, and the judgement a buyer ultimately makes.

He called attention to three gaps between those stages.

The proof gap is the distance between a company’s claim and credible external corroboration. If a business says onboarding is straightforward, what do customers, practitioners or independent reviewers actually report about onboarding?

The citation gap asks whether existing external evidence appears in the AI answers being investigated. Credible reviews or practitioner explanations might exist but not show up in a particular sample of answers. That is a discoverability observation, not evidence that a platform owes the organisation a citation.

The trust gap is harder: even if an answer surfaces the claim and evidence, does a buyer find the information persuasive, relevant and credible enough to affect their judgement? An AI monitoring dashboard cannot answer that question alone. It calls for actual customer research, sales insight or appropriately designed user testing.

Indig’s concluding advice was that brands do not have to be everywhere; they need to close the gaps that matter to them. That is the useful constraint. The framework is his. What follows is how I would turn it into a bounded editorial investigation rather than another permanent content workstream.

As I argued in When search becomes recommendation, I wouldn’t describe AI systems as trusting anyone; the work is improving the information people rely on.

Start with a claim, not a channel

A predictable response to the third-party-source argument would be to order more reviews, secure more coverage, commission YouTube content and ask the social team to spend more time in communities.

That is how an apparently strategic insight becomes four disconnected programmes, each with its own metrics and nobody responsible for the underlying question.

I would start smaller. Select five consequential claims the organisation makes to a defined audience. They might concern ease of adoption, suitability for a particular type of customer, accuracy, support, reliability or the measurable outcome of a process change. The best candidates are claims that appear on the homepage, pricing page or sales deck and that are either used most in sales conversations or challenged most in lost deals. Put them in the exact words the business uses, rather than translating them into vague themes.

If five is too many for the team you have, start with one claim, two AI tools, an ordinary web search and an afternoon. That is enough to find out whether the method is worth extending.

Then ask what an intelligent sceptic would need to establish whether each one holds. Not every claim needs the same kind of proof. A technical compatibility statement should be checked against product documentation and a demonstrated implementation. A statement about helpful support needs customer experience over time. A cost-saving claim needs a clear basis of calculation, context and limitations.

The important move is to translate the marketing sentence into a question that can be investigated.

Brand claim	Question a buyer might ask	Useful corroboration
Easy to set up	What does implementation involve for an organisation like mine?	Walkthrough, implementation experience, current onboarding documentation
Saves time	Which tasks became quicker, by how much and under what conditions?	Attributable customer example, verified before-and-after process
Built for my industry	Does it handle the exceptions and constraints common to my sector?	Specialist practitioner testimony, relevant product evidence
Better than alternatives	How does it compare on the features that matter to my situation?	Fair comparison, current specifications, independent evaluation
Trusted support	What happens when a customer encounters a real problem?	Documented resolution processes, representative customer feedback

This is an illustrative working table, not evidence that any named company makes or satisfies these claims. In a live audit, the product and audience constraints must come first.

Look at the answers before making the distribution plan

Once the questions are clear, inspect a manageable set of relevant AI answers. Use the models and markets your audience actually uses where there is evidence for that choice, and preserve enough context to reproduce the observation: prompt, engine, date, market and any retrieval or search mode available.

For each answer, separate four observations.

Portrayal: What is said about the organisation? Is it described accurately, and in which customer context?
Mention: Is the brand named or recommended? Is the inclusion favourable, qualified or unsuitable for the request?
Citation: Which specific pages or videos are cited? Do they genuinely support the answer, and who owns them?
Buyer relevance: Would this information help a person with the stated problem make a sounder decision?

Do not treat a citation as endorsement. An answer can cite a review criticising a product. It may use a publisher as a source without mentioning the publisher by name. A brand can appear in a list because it is familiar while the answer offers almost no evidence for its suitability.

I would also search outside AI tools. What does an ordinary web search reveal about the claim? What do customers actually say in relevant reviews? Are practitioners discussing it in professional communities? Is there a demonstration from somebody who has used the product? Are authoritative sources saying something that the company’s own page overlooks?

Then list the information competitors for the question: the sources an answer relies on to explain the problem, whether or not they sell anything. For each one, note what it does better than you. A professional body may be the right authority on the regulation, and the useful response is to link to it and stop competing. A publisher’s explainer may simply be clearer than your guide, which is an editorial problem you can fix. A practitioner’s video may show the process honestly where your page only describes it, and the right move may be to offer that practitioner better information or access rather than to outrank them. Treating every such source as a rival misses the more useful question: which of them should your evidence be feeding?

The Semrush UK findings are a reminder to repeat this in the actual market and category, rather than assuming that a source pattern seen in a US retail benchmark applies to UK professional services. AI results are volatile; a single prompt run is a useful observation, not a stable estimate of market position.

A worked example: the setup claim that falls apart

Suppose a fictional B2B software company says that implementation is simple. Its website supports the claim with an attractive product tour and a case study describing a smooth launch.

A small audit finds two independent reviews from customers describing a difficult data migration, one detailed practitioner video showing the configuration process, and a recent article comparing the implementation requirements of several products.

There are at least three ways to respond, and they lead to very different work.

The first possibility is that the criticism is accurate but applies to a particular use case. The company should not try to make it disappear. It may need to explain which customers can use the straightforward route and which need specialist migration support. A more honest answer could be commercially useful, even if it is less flattering.

The second possibility is that the reviews reflect an older product or process. Then the first job is to verify what actually changed. A newer customer implementation or a current demonstration might help people assess the product fairly. The correct response to an old review is not to pressure its author into changing it; it is to make accurate current information available and respond transparently where the platform permits.

The third possibility is that the negative evidence reveals an unresolved product problem. In that case, the best editorial recommendation may be to refer the issue to product and customer-success teams, not publish another article. Content cannot make an implementation experience easier by describing it more enthusiastically.

Only after that diagnosis would I worry about whether an AI system cites the current demonstration, an independent reviewer or the product page. The evidence has to be worth finding before discoverability is the main problem.

This is the point I also made in GEO sentiment needs more than a score: an unfavourable portrayal may be misleading, outdated or entirely fair. The task is to find out which before trying to change the answer.

Before deciding the problem is discoverability, check three things. Does the supporting evidence exist in public, in text a crawler can read, rather than in a PDF, a gated asset or a video with no transcript? Does it appear in an ordinary web search for the buyer’s question? In his published research, Indig found that ranking in conventional search was the factor most strongly correlated with AI citation, so evidence that can’t be found there is unlikely to be cited. And is it specific enough to answer the question as asked? If the evidence fails the first two checks, it is a discoverability problem for the web or SEO team. If it fails the third, or doesn’t exist, it is a proof gap, and technical work won’t close it.

Choose the platform by the proof the buyer needs

Indig offered four buyer questions at Spotlight that I think are particularly practical. They do not dictate a universal channel strategy; they help identify where a particular kind of evidence might be credible.

“Will it work for me?” often leads to reviews and customer accounts from people with comparable needs. In a complex B2B decision, similarity of situation matters more than the total number of stars. A review from a five-person practice does not establish suitability for a multinational finance team.

“Can I see it work?” suggests demonstrations and video. The value is not that YouTube is inherently an AI-ranking tactic, but that a real walkthrough may reveal behaviour a polished product claim cannot. I explore the video-specific audit in Is YouTube the overlooked channel in your GEO strategy?.

“What do peers say?” points towards practitioner communities, professional conversations and credible user experience. Engagement should add information or answer a genuine question, not impersonate a customer, seed covert promotional posts or treat Reddit threads as assets to capture.

“How does it compare?” may call for independent publishers, specialist reviewers or evidence-based comparison resources. A publisher is under no obligation to repeat the brand’s preferred positioning. Provide accurate specifications, transparent limitations and access to evidence; accept that genuine evaluation may conclude a competitor is better in some situations.

I would avoid interpreting this as a command to maintain a presence on every platform. Review sites, video, communities and publishers are potential homes for different forms of evidence. The priority depends on the actual buyer question and on which sources show up in the relevant information environment.

Some claims may be better supported by official guidance or independent testing than by any of those four. Source relevance should win over framework neatness.

Outside proof has to be independent enough to mean something

There is a tension at the centre of this approach. Marketers would like other people to substantiate their claims, but evidence becomes less persuasive when those other people are effectively repeating a brief supplied by the brand.

A sponsored comparison can still be useful if its commercial relationship and methodology are visible. A customer story can still be important first-hand testimony even when the customer approved the final copy. A company can supply a journalist with accurate data without controlling the resulting judgement.

But none of these is equivalent to an independently conducted review. We need to preserve that distinction.

My test would be to ask: could this source reasonably disagree with the organisation and still publish its conclusion? If not, its contribution may be useful distribution, but it is weaker as independent corroboration.

This also means looking for evidence that contradicts the claim. An organisation that only audits favourable reviews is not checking whether its assertions hold. It is collecting testimonials. A disciplined audit records what was said, the source’s incentives, when the experience occurred, what can be verified and what remains unknown.

It also means being careful about how outside evidence is encouraged. Asking only satisfied customers for reviews, offering incentives without disclosure, or nudging critics to revise what they wrote all produce corroboration that looks independent and isn’t. In the UK, the CMA’s fake reviews guidance under the Digital Markets, Competition and Consumers Act treats fake reviews, concealed incentivised reviews and misleading selection of reviews as banned practices. Those rules are written for consumer markets, but they’re a sensible standard for B2B teams too: if a tactic would be unlawful with consumers, it’s no basis for an evidence strategy with businesses.

The 2025 Edelman–LinkedIn B2B Thought Leadership Impact Report, a survey of nearly 2,000 professionals, found that “hidden buyers” in a buying group actively seek out and evaluate thought leadership, much as the main decision-maker does. It’s worth being clear about which way that cuts: the material in question is mostly produced by vendors, so the report is a reminder that a company’s own content can carry real weight when it offers genuine insight. What it adds here is that credibility has to hold up with people the vendor never speaks to directly. It says nothing about AI citations.

Where this argument runs out

There is a strong case against all of this. In many B2B decisions, the proof that actually settles a purchase is private: a reference call with a customer the buyer trusts, a trial on their own data, a security review, a contract with service levels. None of it is visible to an AI system or a web search, and none of it needs to be. For a specialist product with a few hundred customers, there may never be enough independent public material to corroborate anything, and demanding it could push a team towards manufacturing the appearance of independence.

I think that is right as far as it goes, and it changes what the audit is for. Public corroboration matters most before a buyer is willing to talk to sales: the shortlist, the AI answer, the colleague checking a vendor before a meeting. It matters less once private proof is available. So for each claim, the audit should ask where in the decision it is actually tested. A claim settled in a reference call may need no public proof programme at all. A claim that decides whether you make the shortlist probably does.

Turn the findings into a decision, not another dashboard

I would keep one small evidence map for the initial five questions:

Record	What to capture
Claim and audience	Exact statement; customer situation in which it is supposed to hold; where in the decision it is tested (shortlist, sales conversation, trial, reference call)
Buyer question	What someone would need answered before relying on that statement
Owned evidence	Product documentation, research, customer account, verified data
Outside evidence	Source, link, date, independence, what it supports or challenges
AI observation	Prompt, model, market, date, portrayal, citations and mentions
Gap	Proof gap, citation gap, trust gap — or no meaningful gap
Next action	Correct, investigate, acquire evidence, clarify, route to owner, retain or stop
Success check	What would count as a better answer or a better-informed decision?

This is not a new score to sell as scientific certainty. It is an editorial working record. If it becomes too difficult to maintain, make it smaller.

At the end of the first pass, I would choose one or two actions. Perhaps the product page needs to state implementation limitations more clearly. Perhaps a customer interview should establish what actually changed during migration. Perhaps an independent reviewer needs accurate current specifications. Perhaps the reputational problem is real and belongs with customer-success leadership.

Then test whether the action improved something observable. Better factual coverage on a page, a more specific customer example, a corrected source, a practitioner who can now explain the process, or a repeatable improvement in the answers sampled over time are possible outcomes. Record which one you actually achieved.

Re-sample the same prompts four to eight weeks after the action, and treat a single run as an observation rather than a result. Do not infer a sales impact from a citation movement alone. A change in AI answers after publication can reflect model variation, new sources or other activity. To understand buyer judgement, ask buyers: add one question to win/loss interviews about what they read or asked before talking to sales, or show three to five customers the AI answer and ask what they would conclude from it. If that evidence is unavailable, name the missing evidence rather than turning the hypothesis into a result.

An editorial team cannot own the whole evidence environment

The other awkward conclusion from Spotlight is organisational. Kevin Indig’s trust-layer approach and the cross-channel discussion involving Crystal Carter and Eli Goodman both point beyond a conventional content production team.

Editorial teams can identify the question, trace the source, interview the person with first-hand knowledge, test the accuracy of a claim and make a stronger public explanation. They cannot personally fix product limitations, own every review platform, run all partnerships, conduct independent journalism on demand or direct how an AI company chooses citations.

A useful diagnosis therefore names an owner. Editorial may own an article or interview. Product owns the actual capability. Customer success can investigate repeated problems. PR and partnerships may be responsible for relationships with external publishers. Video, social and technical teams own their particular implementations. Someone senior has to resolve conflicts when the desired brand story diverges from customer experience.

Otherwise an evidence audit becomes another content backlog. The information problem is noticed by the only team willing to investigate it, which is then expected to solve everyone else’s part of the work too.

Your website still matters. It just cannot be the only witness.

A good owned website remains essential. It should explain the product accurately, make important evidence accessible, disclose limitations, keep facts current and help people understand whether the offer suits them. Technical SEO and crawlability still have work to do.

But some claims are inherently about experience outside the organisation’s control. You cannot prove that customers find your support helpful merely by publishing a page saying they do. You cannot demonstrate that practitioners recommend a workflow by writing the recommendation on their behalf. You cannot settle an independent product comparison by declaring yourself the winner.

In the Semrush/Datos panel, Eli Goodman illustrated this with a Capital on Tap example: a composite of observed journey patterns, in which research moved between AI tools, video, forums and other sources before an application on the website. It wasn’t one buyer’s journey, and it came from a company that sells journey analysis, so I’d treat it as an illustration rather than evidence of how most buyers behave. Its use is to challenge a last-click assumption: a decision may already be taking shape somewhere the brand doesn’t own.

That is why the first two questions I would put to a B2B content team are not “How many new pages should we publish?” and “Which platforms should we optimise?” They are “Which of our claims matter most to buyers?” and “Who besides us can show that those claims are fair?”

If the answer is nobody, there may be a genuine evidence gap. If the answer is a credible critic, there may be something more important to fix than discoverability. If the answer is already a customer, practitioner or independent source, the next step may be to help that knowledge become accessible without turning it into another piece of branded assertion.

Sometimes the most useful finding will be that a claim should be made more narrowly. That isn’t a failure of the audit. A claim that holds up when someone else checks it is worth more than a broader one that only your website is prepared to make.

Source and method note: This article develops arguments already made in When search becomes recommendation, evidence becomes the interface and Content teams need an evidence layer. The conference interpretation draws on Kevin Indig’s trust-layer presentation and the Semrush AI visibility panel at Spotlight London, 13 October 2026, cross-checked with the published Semrush UK AI Visibility Index findings. The worked software example and audit template are illustrative editorial recommendations, not measured outcomes or proven AI-ranking factors. The reported Capital on Tap example is explicitly composite, not an individual buyer case study. Semrush and Datos, whose research is cited here, sell AI-visibility and journey-analysis products.