A buyer can land in ChatGPT, see your competitor named first, and still show up on your site three days later. By then, the shortlist was already shaped somewhere you could not see in your usual dashboard.
Attribution for AI search needs three layers: directly measurable signals, correlational signals, and things you should not claim at all. For UK marketing teams, the right board-level question is not “did AI drive the deal,” but “did our answer share rise before the pipeline moved, and can we prove the direction with clean evidence?”
What UK teams can measure right now
The measurable layer is small but useful. You can track AI referral traffic, self-reported discovery, and share of answers on a fixed set of buying questions.
If you run a B2B SaaS site in the UK, that means checking whether visitors arrive from sources such as chatgpt.com or perplexity.ai, asking buyers how they heard about you, and monitoring whether ChatGPT, Claude, Gemini, or Google AI Overviews name you on the prompts that matter. Ofcom notes that about 30% of searches now show AI overviews, 53% of UK adults say they often see them, and ChatGPT had 1.8 billion UK visits in the first eight months of 2025 versus 368 million in the same period of 2024, so this is no longer fringe behavior: Ofcom’s 2025 UK online habits report.
AI referral traffic is measurable, but noisy
Referral traffic from AI assistants is the cleanest signal you have today. OpenAI says publishers who allow OAI-SearchBot can track referral traffic from ChatGPT in analytics, and ChatGPT automatically includes utm_source=chatgpt.com in referral URLs, which makes this a real source line rather than a guess: OpenAI publishers and developers FAQ.
That said, referral traffic undercounts reality because assistants strip context, some apps collapse or mask the referrer, and some users copy links into a browser after seeing an answer. A PMM at a B2B software business should treat AI referrals as a floor, not a total, and annotate the report with the source names that do appear, not with a fantasy total.
- What to log: source domain, landing page, query or prompt theme if available, and assisted conversion path.
- What to watch: spikes from chatgpt.com, perplexity.ai, and any assistant-branded source that shows up in analytics.
- What to avoid: rolling all “direct” traffic into AI influence without proof.
Self-reported discovery still matters
“How did you hear about us?” looks old-fashioned until AI changes the first touch. For a procurement tech team or an HR tech team, that question catches the cases where a buyer saw your name in an assistant, then came back later through branded search or a direct visit.
Use a short list, not an open text field alone: ChatGPT, Google AI Overviews, Perplexity, Claude, Gemini, colleague recommendation, review site, LinkedIn, and other. The point is not perfect memory. The point is to separate assistant-influenced demand from everything else before the lead hits the CRM.
Share of answers on tracked buying questions
Answer share is the cleanest upstream metric for AI search visibility. If your brand appears in four out of ten tracked prompts about “best payroll software for UK contractors” or “which vertical SaaS platform handles multi-entity billing,” that trend matters even when traffic is flat.
Ofcom’s distinction between “GenAI search” as chatbots and AI summaries matters here because the touchpoints differ by interface: answer share in ChatGPT is a different measurement problem from visibility in Google AI Overviews. The section in this article’s field notes shows the same pattern across multiple buyer questions, and the practical lesson is simple, if the answer layer shifts first, the traffic layer may follow later.
| Signal | What it tells you | How clean it is | Good for board reporting? |
|---|---|---|---|
| AI referral traffic | Assistant-sourced visits that reached your site | Medium, because of referrer loss and dark traffic | Yes, with caveats |
| Self-reported discovery | Whether a buyer remembers an AI touch | Medium, because memory is imperfect | Yes, as a directional overlay |
| Share of answers | How often you are named on tracked prompts | High, if the prompt set is stable | Yes, as a leading indicator |
| Per-answer impression counts | Nothing you can verify today | Not available | No |
What is correlational, not causal
Correlational signals help you see movement, but they do not prove cause. If answer share rises and inbound leads later rise, you can say the relationship is plausible and worth tracking, not that the assistant caused revenue.
That distinction matters because UK teams are budget-scrutinized. A demand gen lead reporting to a CFO should not present AI answer growth as pipeline attribution. Present it as a leading indicator, then line it up next to downstream outcomes and let the trend speak for itself.
Inbound lift after answer-share lift
This is one of the most practical correlational reads. If your brand starts appearing more often in assistant answers on a stable set of buyer questions, and inbound visits or assisted conversions lift a few weeks later, that is a pattern worth believing.
My view: this is the right way to report AI search optimization internally. Treat answer share the way SEO teams once treated rank tracking, as an upstream signal that deserves its own trend line beside pipeline, not as a replacement for commercial metrics.
Branded search rising after assistants name you
Branded search is another plausible correlate. When buyers see your name in ChatGPT, Claude, Gemini, or Google AI Overviews, they may later search your brand directly, especially in categories with long evaluation cycles such as cyber security, HR tech, martech, or legal tech.
Pew’s July 2025 study found that when Google shows an AI summary, users are less likely to click through to other websites, which supports the idea that visibility inside the answer layer changes the path before the visit happens: Pew Research Center. That is one reason branded search often moves before your traffic chart does.
Country and category change the read
UK buyers do not behave like US buyers in every category, and AI answers do not behave like classic search results. A payments infrastructure vendor serving London-based fintechs may see assistant exposure in shortlist questions long before it sees volume in branded search.
For UK-specific planning, anchor your work in AI search optimization in the United Kingdom and then compare the pattern against your other markets. A brand can be visible in the UK, invisible in India, and only partially named in the US, even when the same product page is live everywhere.
What you should never claim
Do not claim impression counts inside assistants today. Do not claim per-answer view data. Do not claim a precise share of exposed users unless the platform itself gives you a public, auditable source for that number.
The European Commission’s AI Act transparency framework is useful here because it focuses on disclosure, labeling, and transparency obligations, not on handing marketers per-answer logs or impression ledgers: European Commission AI Act transparency guidance. That is the boundary. Anything beyond it is a guess dressed up as measurement.
The trap of synthetic precision
Boards hate hand-waving, but they hate fake precision more. A slide that says “our brand received 18,742 AI impressions” without a source chain is weaker than a slide that says “our answer share rose across 12 tracked buying prompts, and referral traffic from assistant domains moved in the same direction.”
Pew also found broad exposure and mixed trust, with 65% of U.S. Adults saying they at least sometimes encounter AI summaries in search results, but only 6% saying they trust the information a lot. That is a reminder to interpret visibility cautiously, because seeing an answer does not mean the buyer accepts it: Pew Research Center.
What not to put on a board slide
- Per-answer impressions: unavailable today, and usually inferred from nothing.
- Total assistant views: not exposed in a marketer-friendly way.
- Single-prompt screenshots: variance, not trend.
- Revenue certainty: impossible to prove from answer visibility alone.
How to read answer share like rank tracking
Answer share should sit next to pipeline the way rank tracking once sat next to organic traffic. It is an upstream trend line, not the commercial outcome itself.
That framing is useful for a PMM at a B2B SaaS business because it gives the team a disciplined way to talk about AI search optimization without promising what cannot be proven. One chart shows whether the brand is getting named more often. The other shows whether downstream demand moved afterward.
Build a stable prompt set
Choose 10 to 20 prompts that mirror real buying questions, then keep them fixed for a quarter. Use category, comparison, and constraint prompts, not broad curiosity prompts.
For example, if you sell vertical SaaS, use prompts around shortlist behavior, pricing, implementation effort, and UK fit. If you sell cybersecurity software, use prompts around compliance, team size, deployment, and reporting. If you sell HR tech, use prompts around payroll complexity, region, and integration depth. Stable prompts make the trend line believable.
Track movement in the right order
- Answer share first: check whether your brand is named, recommended, or compared.
- Referral traffic second: see whether assistant domains send actual visits.
- Branded search third: watch whether more people search your name afterward.
- Pipeline last: read it as downstream context, not proof of attribution.
McKinsey describes AI-powered search as a new front door to the internet, and it also notes that remaining clicks from traditional search are more likely to come from better-informed consumers later in the funnel: McKinsey. That fits what many UK teams are already seeing, a buyer has the conversation elsewhere, then arrives with a narrower set of options.
Data-quality traps and how to handle them
The biggest attribution mistakes are boring ones. Referrer stripping, dark traffic, and messy lead forms create fake certainty, then the reporting deck gets built on top of it.
If you run content or demand gen, fix the plumbing before you claim insight. The goal is not perfection. The goal is a clean enough signal that the board can trust the direction.
Referrer stripping
Some AI surfaces pass referrers cleanly, and some do not. If the referrer drops, the visit may appear as direct traffic, email, or something else entirely.
Practical mitigation: tag AI-recommended links where you control the destination, watch for source domains that are explicitly returned, and compare spikes against published content updates or campaign launches before assigning credit. OpenAI’s note that generated links can transmit user context to destination sites is another reason to keep privacy review in the loop: OpenAI help article on generated links.
Dark traffic
Dark traffic is the chunk you never see because the source is hidden, shortened, copied, or passed through private apps. That is normal in AI search and in messaging channels.
Mitigation is mostly process. Use self-reported discovery, keep branded search separated from generic search, and compare trends over several weeks instead of reacting to a single spike. Dark traffic is why answer share matters. It gives you a signal before the visit becomes visible.
Bad baselines
A baseline taken after a major content refresh is a weak baseline. So is a baseline that mixes regions, engines, and prompt intent into one line.
Keep the UK separate where the market matters, and keep ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews distinct where the interface matters. Ofcom’s definition of GenAI search is helpful precisely because it reminds you that chatbots and summaries are not the same measurement object: Ofcom’s GenAI search qualitative research.
What the panel of buyer prompts shows
The panel shows two things that matter for attribution. First, buyers often encounter AI answers before they hit your site. Second, the same brand can be named in one engine and omitted in another, which means single-engine reporting is too thin.
That matters for B2B SaaS teams because category behavior changes by buyer task. A martech buyer asking for comparison language may see different names than a data and analytics buyer asking for implementation depth, and the metrics should reflect that split.
Observation one: the answer layer is where the first measurable movement often appears, not the traffic layer. That is why answer share belongs on the board slide even when referral traffic is still quiet.
Observation two: engine differences are real. ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews do not surface the same evidence patterns, so a single “AI visibility” number hides too much.
A 30-minute measurement setup you can run now
You can build a useful AI search measurement sheet in under 30 minutes. You do not need a platform to start, just discipline.
- Pick 12 prompts: four category prompts, four comparison prompts, and four constraint prompts tied to your product.
- Test four surfaces: run the same prompts in ChatGPT, Claude, Gemini, and Google AI Overviews.
- Score each answer: 0 if unnamed, 1 if mentioned, 2 if recommended or compared favorably.
- Log one evidence note: record the page, review, or citation that seems to support each mention.
- Repeat weekly: keep the prompt set fixed for a month.
This gives you a trend line you can defend, and Cited can automate the scaled version with weekly re-checks, diagnosis, and editable fix drafts through why Cited or a free start at /start.
How I would report this to a board
I would not report AI search attribution as a miracle metric. I would report it as a visibility system with measured layers.
The slide would have three lines: answer share on tracked buying prompts, AI referral traffic from known assistant domains, and branded search or assisted pipeline movement after visibility improves. That is enough to show whether the market is hearing you earlier, even if the final sale still needs human validation.
For a UK marketing team, that is the practical answer to “what can we actually measure?” Measure the signals that exist, label the ones that only correlate, and refuse the ones that do not exist yet. That discipline is what keeps AI search optimization credible inside a budget review, and it is the same reason teams use generative engine optimization and answer engine optimization as operating terms rather than slogans.
If you want a deeper view of how to turn those measurements into a working operating model, the broader platform approach is laid out in our AI search visibility audit guide and the GEO platform comparison. The point is not to chase a prettier dashboard. The point is to know whether buyers are hearing your name before they decide you are on the list.
Frequently asked questions
What can we actually measure in AI search attribution?
You can measure AI referral traffic, self-reported discovery, and share of answers on a fixed prompt set. Those are the only layers in this article that are directly measurable enough to defend. Everything else should be treated as correlational or not claimed at all.
How do I know if ChatGPT or Google AI Overviews sent a visit?
Check your analytics for assistant-branded source domains such as chatgpt.com or perplexity.ai, and look for UTM tags where they appear. OpenAI says ChatGPT can include utm_source=chatgpt.com in referral URLs, but some traffic still gets masked or lost.
How is answer engine optimization (AEO) measured?
By answer share: of the AI answers generated for your tracked buying questions, the percentage that names your brand. Read it like rank tracking for answer engine optimization - a leading indicator with its own trend line, sitting upstream of the traffic and pipeline numbers your board already knows. AI search analytics tools automate the collection; the metric definition stays the same.
Can I report AI search as pipeline attribution to my CFO?
Not directly. The article argues that answer share and AI referral traffic are leading indicators, while pipeline is downstream context. A board slide should show movement in those layers, not pretend you have per-answer revenue attribution.
What is the best way to measure AI search in the UK?
Keep the UK separate in your baseline, use stable prompts that match real buying questions, and compare ChatGPT, Claude, Gemini, Perplexity, and Google AI Overviews separately. UK behavior and engine behavior both change the read.