What honest AI recommendation measurement requires and what it cannot responsibly claim.
Your dashboard may show the buyer’s arrival. It still cannot show the full AI evaluation that helped shape whether the buyer arrived at all.
That line from the last issue struck a nerve. Reader after reader came back to the same concern: part of the buying journey may now happen before conventional analytics register anything. If that evaluation is real and consequential, the next question is unavoidable.
Can we actually measure whether a business is earning recommendation from AI systems – honestly?
TL;DR
AI is already part of B2B research and evaluation, but adoption does not eliminate the buyer’s need to validate what the system says.
One favorable answer is an observation, not a measurement. It shows what one system produced in one context at one moment.
Credible measurement requires questions fixed in advance, repeated observations, defined systems and buyer contexts, representation analysis, a true baseline, and claims bounded by normal variability.
Recommendation Marketing should refuse guarantees, permanent scores, single-cause attribution, and any conclusion the evidence cannot responsibly carry.
The short answer is that we can observe and compare a meaningful slice of AI-mediated evaluation. We can ask questions that real buyers ask, record which businesses appear, examine how they are represented, note what evidence is cited, and repeat the process over time. But the method matters because a measurement practice should be capable of producing an answer we did not want.
The measurement gap is now a business problem
The need for disciplined measurement is no longer theoretical. Buyers are already using generative AI to discover, compare, and evaluate businesses, but the available research also shows that they do not simply accept every answer at face value.
Forrester reported that 89% of the B2B buyers in its 2024 Buyers’ Journey Survey were using generative AI in at least one area of their purchasing process. Among buyers who used it, 87% agreed that it helped create a better business outcome for their organization. Those findings show why AI-mediated evaluation deserves attention. They do not tell an individual business whether it is being represented accurately or whether a particular recommendation caused an outcome.
Source: Forrester, The Future of B2B Buying Will Come Slowly and Then All at Once
Gartner’s more recent survey of 645 B2B buyers, conducted in August and September 2025, found that 45% had used generative AI, primarily to gather information about vendors and products. Sixty-nine percent preferred to validate AI-generated insights with sales representatives, while 51% said they were more likely to encounter misleading information from generative AI. Adoption and skepticism can exist at the same time.
Source: Gartner, Survey Finds 69 Percent of B2B Buyers Turn to Sales Reps to Validate AI Generated Insights
The Forrester and Gartner percentages should not be compared as a trend. The studies used different samples, questions, definitions, and time periods. Together, they support a narrower conclusion: generative AI has become a material part of B2B buying, and buyers still need validation. That makes honest observation more important, not less.
The claim agencies have always made
I have spent more than thirty years in this industry, so I can say this from the inside: every agency has claimed that its work produces results. The claim has often been sincere. It has rarely been proven all the way through to the outcome the client actually wanted.
Instead, we measured activity that stood in for outcome. Impressions served. Rankings achieved. Clicks delivered. Engagement generated. Those metrics are real and useful. They show that something happened. They do not, by themselves, prove that the right buyer trusted the business, included it in the serious set, or chose it because of the work.
The industry was not necessarily being dishonest. The decisive part of the process was difficult to observe. We could see the media, the visit, and the form submission. We could not sit inside the buyer’s deliberation and watch a particular piece of marketing earn confidence.
AI recommendation does not make the entire decision visible. It does make one consequential layer more observable. We can now ask a system the question a buyer might ask and examine the response that comes back.
That is progress. It is not yet proof.
Observable does not mean measured
Justin Hu raised the right challenge on the first issue of this newsletter: which query would actually put you in the answer this week, and have you checked whether it still says you tomorrow?
Ask the same question again and the answer may change. Ask it on another system and it may change again. Add a location, company size, risk tolerance, or use case and the field may shift. The model, retrieved sources, phrasing, context, and timing can all affect what appears.
One favorable answer is not measurement. It is a testimonial from a machine, and it deserves the same skepticism as any other testimonial.
A screenshot of a flattering answer is an anecdote. A dashboard that checked once is a snapshot of a moving target. Neither can responsibly answer the questions a skeptical buyer or CFO should ask: compared with what, measured how, across which conditions, and would the conclusion hold tomorrow?
Six requirements of honest measurement
A credible method must be structured so that it can be wrong and disciplined enough to reveal when it is. Six requirements follow from that principle.
1 Fix the questions before measuring
Choose the buyer questions that matter before observing the answers, then hold the core set constant. If questions can be changed after results appear, the method can always be made to flatter. A fixed question set makes before and after comparable rather than merely narratable.
2 Repeat the observation
A single run cannot tell you what is stable. Repeated runs help distinguish durable patterns from normal answer variability. Variability is not a defect in the method; failing to account for it is.
3 Span systems and buyer contexts
Buyers do not all use the same platform or describe their needs the same way. A single system and a single phrasing measure only that condition. The meaningful finding is the pattern across defined systems, phrasings, and buyer contexts.
4 Record representation not just appearance
Being named is not automatically a win. A business can appear while being described inaccurately, associated with the wrong expertise, or supported by outdated sources. Measurement should capture whether the business appears, how it is characterized, what evidence is cited, and whether the representation is true.
5 Establish the baseline before improvement work
A baseline captured after changes begin is not a baseline. The before picture must be documented and frozen while it is still the before picture. That can be uncomfortable. A real baseline may expose absence, inconsistency, or misunderstanding. That is precisely why it is useful.
6 Report movement inside honest bounds
Any before-and-after comparison must respect the variability found through repetition. A change smaller than the system’s ordinary movement should not be presented as meaningful progress. Honest reporting distinguishes observed movement from a conclusion the evidence cannot yet support.
What honest measurement refuses to claim
The same discipline that makes measurement credible also establishes boundaries. Four matter most.
No guaranteed inclusion. No agency controls what an AI system will say or whom it will name.
No single-cause attribution. Systems, competitors, sources, and market conditions change alongside the work.
No permanence. Standing can be earned, weakened, or lost as evidence ages and conditions change.
No universal formula. Platforms differ, buyer questions differ, and a single score cannot explain every evaluation environment.
Confidence is easy to publish. Deserved confidence requires a method that could have embarrassed you and did not hide the result.
What this changes for marketing
An industry that cannot demonstrate outcomes tends to sell activity. Clients pay for hours, deliverables, media, and production, then try to infer whether those inputs changed consideration. That gap is one reason many business leaders carry quiet skepticism about marketing even when they continue to invest in it.
A more observable evaluation layer creates the possibility of greater accountability. It does not eliminate attribution problems or reveal every influence on a buyer. It does allow us to compare how a business is understood and represented across a governed set of questions before and after deliberate work.
That is not perfect measurement. It is more honest measurement and it is enough to change the conversation from what we produced to what the market can now understand, verify, and recommend.
Right Angle will be measured by this standard
It would be easy for me to describe this method and apply it only to other businesses. That would miss the point.
Right Angle will establish its own governed baseline: a fixed set of questions our buyers ask, observed repeatedly across defined systems and contexts, with appearance, representation, evidence, and variability recorded before improvement work begins.
The baseline is not ready for publication today. I will not fill that space with an early screenshot or characterize movement before there is a documented before picture. When the baseline is ready, I will publish what it finds including anything I would have preferred it not find.
If you ran the audit from last week, you already have a first observation. Run it again. Record what held, what changed, and whether your business was represented accurately. Then share the result with me. The pattern across readers will be the subject of the next issue.
Frequently asked questions
Can AI recommendation be measured?
A repeatable slice of AI-mediated evaluation can be observed and compared. A credible method can record whether a business appears, how it is represented, which sources are cited, and how those results vary across defined questions, systems, contexts, and dates. It cannot measure the buyer’s complete deliberation or prove that an AI answer caused a purchase.
Why is one favorable AI answer not proof?
AI answers can change with the platform, model, phrasing, context, retrieved sources, and timing. One favorable result documents what happened once. It does not establish stability, improvement, buyer trust, or causation.
What should an AI visibility baseline record?
The baseline should record the fixed buyer question, platform, date, relevant context, whether the business appeared, how it was described, which sources were cited, whether the representation was accurate, and how the result varied across repeated observations. The baseline must be captured before improvement work begins.
Can an agency guarantee that an AI system will recommend a business?
No. An agency can improve the clarity, credibility, evidence, consistency, and technical accessibility available to AI systems and buyers. It cannot control a platform’s outputs or responsibly guarantee inclusion, position, permanence, or revenue.
How do you know whether a change is meaningful?
First establish the ordinary variation produced by repeated observations under defined conditions. A later change should be described as meaningful only when it is large and consistent enough to exceed that normal movement and when the method rules out simpler explanations. Until then, report the change as an observation rather than proof of improvement.
Can a business perform this measurement without a specialist platform?
A business can begin manually by fixing a small set of buyer questions, recording the conditions of each test, repeating the observations, and comparing appearance, representation, and citations. A specialist platform can improve consistency, scale, and recordkeeping, but tooling does not remove the need for sound questions, human review, and bounded claims.
If you would rather not wait to understand how your business currently appears across the questions that matter, Right Angle combines Answer Engine Optimization with Sightline to examine visibility, representation, citations, and competitive presence.
Coming next: what readers found when they repeated the audit – where the answers held, where they changed, and what those differences teach us about measuring a moving system.
Visibility + Engagement = Opportunity
