Expert Mode: Why a Citation Isn’t a Recommendation with Kelvin Çobanaj from ZeroRank

This article is based on a Q&A with ZeroRank CEO and Co-Founder Kelvin Çobanaj and Greg Kihlström for The Agile Brand Guide’s Expert Mode series.

Most of us have accepted that AI answers now shape the shortlist a buyer walks into a sales conversation with. The instinct that follows is a familiar one. We want a number for it, the way we’ve had rank and share of voice for search, so it can go in the monthly report and, ideally, go up.

The platforms don’t cooperate. When BrightEdge, which sells its own AI visibility tooling, ran tens of thousands of identical prompts through ChatGPT, Google AI Overviews and Google AI Mode last year, the three disagreed on which brands to recommend 61.9% of the time. Kelvin Çobanaj, CEO and co-founder of ZeroRank, which tracks how AI assistants describe and recommend brands, argues that a single blended score hides exactly what a marketing leader needs to see. In written answers to our questions, he also made a point that’s less convenient for a company in his business: more mentions, on their own, don’t prove much of anything.

AI Visibility Is Several Channels Sharing a Name

We asked Çobanaj whether “AI visibility” is one thing a CMO should manage at all. He didn’t hedge, and he used his own company as the example, which most vendors wouldn’t.

It is several channels serving one goal: getting your brand considered by buyers. I would give a CMO separate results for each platform. Our data shows why. For the same question about “Perplexity tracking tools,” one ChatGPT answer recommended Otterly and Peec, while one Google AI Overview featured Otterly and Orchly. Neither mentioned ZeroRank. One combined score would hide those differences.

A company that sells AI visibility tracking didn’t show up in AI answers about tracking tools, and the two platforms that left it out didn’t even agree on who to include instead.

For the rest of us, the change is in how the report gets built. Instead of one line labeled “AI search,” a team reports by platform, and only against the buying questions that actually precede a purchase in its category. Çobanaj also wants mentions kept apart from recommendations, with links counted separately again. “Being named as a poor fit should never count as the same result as being recommended,” he said. A dashboard that counts every appearance as a win can show steady growth while the answers about your product get worse. If your agency or your generative engine optimization vendor reports a single visibility score, ask them to break it out by platform and by sentiment of the mention before you approve the next phase.

You Control the Inputs, Not the Retrieval

The second question was which of the platforms’ decisions a brand can actually influence. His list of what’s out of reach is longer than his list of what’s in reach.

You can improve the information a platform might find. You cannot choose when it searches, how it rewrites a question, which sources it selects, or how much evidence it shows. Make product details clear and accessible, and correct outdated descriptions where possible… The mistake is reacting to every missing mention by publishing more content. First check whether the product fits the question and whether the available information explains that fit.

Content is the easy thing to buy, and an agency can deliver it on a schedule. That’s exactly why it becomes the default response to a missing mention.

Çobanaj’s order of operations is less satisfying and more useful. Start by asking whether the product genuinely fits the question. If a buyer asks an assistant for the cheapest payroll tool for a five-person company and you sell enterprise payroll, your absence from that answer is correct, and no amount of content should change it. Next, check whether your own public information explains the fit in plain terms. Pricing pages, integration lists and spec sheets that are out of date or locked behind a demo form are a common culprit. Only then does third-party coverage come into it. His own tracking kept finding comparison articles cited in answers that left ZeroRank out. That makes those articles “worth checking,” he said, but “it does not prove that more coverage will solve the problem.”

A Citation Isn’t a Recommendation

Çobanaj’s example came from a remote-hiring platform that used ZeroRank to draft articles on topics like hiring costs and talent comparisons. Only one, a comparison piece, could be confirmed as published. By ZeroRank’s own tracking, the brand’s mention rate in Google’s AI answers across 20 buying questions rose sharply in the period afterward.

A remote-hiring platform used ZeroRank to create eight drafts about hiring costs, talent comparisons, and remote interns. I verified that one comparison article was published… Some answers began describing it as a hiring option. Others cited its articles while recommending competitors. That distinction matters: becoming a source does not automatically mean becoming the recommendation.

The increase is the headline he could have led with. The split is the part a CMO should actually take away. Some of the new appearances were the platform’s own content being used to send buyers to someone else. Plenty of citation optimization work could produce exactly that result and still look like a win in a mentions report.

We also asked him the skeptic’s question directly. AI answers shift with phrasing and session, a volatile metric is an easy way to manufacture urgency, and ZeroRank sells the tracking. When isn’t it worth the budget? He didn’t really answer that. His reply described what ZeroRank does and said repeated sampling of the same question gives teams “a direction to follow, rather than a blind absolute score.” That’s a fair description of what sampling can do, and it’s the right way to read his own case study. One client, one published article, one before-and-after window. His own wording was that “the improvement followed content creation.” Followed, not caused.

So we’ll fill in the part he skipped. If your buyers don’t use AI assistants during research, or the questions in your category rarely return brand names at all, ongoing tracking is an expensive way to confirm that. Run the questions by hand first. Pay for tracking once you’ve seen that the answers in your category name vendors and that they vary enough to need watching.

Show the CFO More Than Mentions

The last question assumed a fixed budget for next quarter and uneven visibility across platforms. His answer was the most practical of the five, and none of it depends on a particular tool.

Start with questions your buyers ask, platforms they use, and problems you can fix at a reasonable cost. An incorrect answer about a required feature may matter more than low visibility for a broad category question. Choose a small project, record the starting point, and agree on what would justify more spending. Show the CFO the cost, changes in recommendations, and results such as qualified inquiries, opportunities, or sales… More mentions alone are not enough to claim more business.

Two parts of that deserve more attention than they usually get. The first is prioritizing by error instead of by gap. An assistant telling buyers your software doesn’t integrate with Salesforce, when it does, is costing you deals today. Low visibility on “best CRM for mid-market” might not be.

The second is the control group. Çobanaj suggested comparing against “products or markets you left unchanged,” which is ordinary incrementality logic, the same standard a CFO would apply to a paid media test. In practice that might mean one product line gets the work and a comparable one doesn’t, both tracked against the same set of buying questions for the same period. Referral data and a “how did you hear about us” field won’t settle the question alone, but alongside a holdout they give the finance team something better than a chart of mentions.

We started with the wish for a single number. Çobanaj’s answers suggest it doesn’t exist in any form worth reporting. What does exist is narrower and more useful: a short list of buying questions, tracked by platform, with recommendations counted separately from mentions. This week, before any new spend, pull the ten questions your sales team hears most often. Run each one through ChatGPT, Perplexity and Google’s AI Mode a few times. Write down which answers are wrong about your product, not just which ones leave you out, and fix those first.

Posted by Agile Brand Guide

Spreading knowledge, one marketing acronym at a time. Content dedicated to all things marketing technology and CX.