The philosophy behind Recommendable
Why you are probably measuring AI Search wrong
Why we built Recommendable: what our research showed about how AI models arrive at a recommendation, the shift in the search landscape, and what we built with it.
In our work for clients we kept getting the same question: why does ChatGPT name my competitor and skip me? That question can be answered with a measurement, and dozens of tools do exactly that. Prompt trackers, let me call them that for now. The question behind it is harder, and it is the only one that really matters: what can you do about it?
To answer that, I went deep into how those models arrive at a recommendation. By now I have read nearly every scientific paper published on the subject, followed dozens of lectures and read everything the providers and the trade press publish. In the same period I tested a great many existing tools. Most of them measure your visibility neatly and leave the rest to you. On a few crucial points they also measure the wrong thing, in my view, because they copy Google's rules onto a completely new playing field.
That research produced thirteen findings. Together they are the reason we built Recommendable, and they explain right away why we did not continue with an existing tool.
Why we built Recommendable
The shift from traffic to the zero-click reality
For years, website traffic was the only holy grail. Marketers got hooked on lines that go up. With AI assistants, however, the conversion or the brand conviction increasingly happens inside the chat interface. The AI summarises your product so well that the user no longer needs to visit your site to be convinced. You are optimising for brand dominance and accuracy in the answer itself, rather than for a click.
Citations are the new vanity metric
Because generating traffic keeps getting harder, many tools now sell citations as the new KPI. But a citation is only a footnote. It proves that the model knows you exist, or used your page as background, not that it puts you forward as the answer. We still measure citations, as an excellent outreach list, but the centre of gravity of the tool is the real KPI: Share of Recommendation. That is where your customers come from.
A probabilistic system needs a different yardstick
With Google you optimised against a system that gave the same answer to the same question. Ask a language model the same question ten times and you get ten variants. One measurement therefore says almost nothing. We measure repeatedly and put next to every figure the band in which the real value lies (95% Wilson). Every number carries a label: Robust, Adequate, Minimal or Interpretation. So you see whether a difference is real or noise, and whether a figure is a measurement or a reading.
AI Search has two modes: live and memory
Live means that a model may search the web, not that it always does. If an engine answers your topic purely from its trained memory, tinkering with your own web pages has little direct effect and online PR (being known) weighs more. We keep the two apart and show per topic in what share of the live answers a search actually happened. That is rather decisive for the tactic you choose.
People do not stop at the first answer
Most tools fire a fixed set of prompts and tally whether you are named in the first response. Nobody talks to an AI assistant that way. There is a follow-up question, an exception, a condition, and only then a choice. We make two things visible:
The decision axes: the conditions on which a model tips its answer, plus whether your site answers them.
Conversations: we let buyer archetypes hold complete conversations with the assistant. You see where you come into view, where you win, where you drop out and the reason the AI itself gives. Every criterion you fail on is something you can repair.
AI thinks in entities, not in keywords
Models do not understand the world through loose fragments of text, but through fixed entities and the relations between them (knowledge graphs). If your brand is not unambiguously defined online, because of poor structured data, a messy company structure or missing external validation, the AI cannot connect the dots. In AI Search, technology is about clarity for the machine, not about keyword density.
Machines follow people, not the other way round
A model forms its picture of you from everything it has found across the internet: reviews, forums, trade press, social. Convince those people and the model follows by itself. That is why Narrative reads everything said publicly about you and your competitors and pulls out the patterns. Site comparison puts what you and your competitors claim on your own pages on one scale, with your ideal customer as the judge. Two brand insights you will not find in a visibility dashboard.
Brand protection matters as much as brand visibility
It is not only about whether you get recommended, but also about the AI not making things up about you. Models regularly hallucinate prices, non-existent features or wrong integrations. Recommendable also works as a radar for this misinformation and reputational damage. Knowing where the model gets your brand wrong is step one in correcting it through PR, FAQ updates or your data structure.
Text is no longer the only language
Today's models are multimodal: they analyse infographics, data visualisations and complete videos as easily as a written article. Does your competitor offer a crystal-clear video explainer or a sharp chart, while you never get past a dry wall of text? Then you lose. The AI always picks the form that conveys the context best and fastest. Optimising for the engine goes far beyond written copy by now.
Without context, every recommendation is a guess
The usual approach: a tool takes your website, has an AI invent a set of prompts and then reverse-engineers the winners. Out comes advice that is the same for every brand in your category. A good consultant starts by asking questions. So does this tool. The intake makes you fill in a profile: who you are, what makes you unique and who you are up against (even per product group). Everything you run afterwards is measured against your specific answers and your market.
AI cannot be hacked (the end of shortcuts)
The traditional SEO world was full of tricks to fool an algorithm. But a language model with billions of parameters is not misled by a few clever keywords or bought links. The model understands context, semantics and the actual reputation of what you offer. You simply cannot hack an AI into recommending a mediocre brand with a poor reputation in the market. That is why you will find no quick fixes or silver bullets in Recommendable, on purpose. Our fundamental conviction is hard but simple: if you want to be structurally recommended as the best choice, you will have to actually be the best choice.
AI Search is brand strategy first
SEO optimises a page for a query. AI Search optimises a brand for a recommendation. Because tricks do not work, getting recommended more often goes straight to your core: your positioning, your proof, your presence beyond your own site and what customers say about you. That is not a lonely technical chore, it takes your whole marketing team. The Action Plan in Recommendable is therefore sorted by impact and effort, with a clear division of tasks for the right team member on every action.
Everything a model finds about you we call your body of evidence: what you claim, what your site proves and what others say about you. A prompt tracker only measures the answer that follows from it. Recommendable measures the evidence itself, what the model has made of it, and only then the answer and the conversation. So when you are absent, you know what it comes down to.
Recommendable does not generate content
That is a deliberate choice. Every tool that produces content at the push of a button delivers mediocre work, and mediocre content in your tone of voice is still mediocre content. Models see through it. We do tell you per page exactly what is missing, why it matters and what it earns you. You export that brief to your copywriter, your agency or your own internal AI assistant. In an internet rapidly flooding with interchangeable AI text, authentic human nuance is your biggest competitive advantage. In the end, quality always beats quantity.
What it adds up to
So Recommendable is not yet another prompt tracker that measures your visibility every day and feeds you false hope about pointless citations. Think of it as the most thorough audit you run periodically in the zero-click era: how do you perform as a brand, where do you drop the ball on the decision axes, and above all, what can you do about it. That produces three layers: strategic insights, tactical plans and operational recommendations. With one goal: influencing the buying decision at the moment it is made.
What is in the platform
- Brand Profile
- Who you are, for whom, against whom, and how you want to be seen. The foundation under everything else.
- Website
- Your site read the way an AI model reads it, sitewide and page by page. More than 40 modules and agents look at technology, content, positioning and how readable you are for AI agents. Two heavier analyses you start yourself per page: the visual assessment of what sits above the fold, and a set of buying-intent prompts that tests whether the page carries the signals such a question turns on.
- Site comparison
- What you and your competitors claim on your own pages, put on one scale and ranked by your ideal customer.
- Narrative
- What the market says about you. Reviews, social, forums and news about you and your competitors, sorted by topic and sentiment.
- Topics
- The demand side of your market: buying moments, the topics under them and the questions customers ask along the way. This is also the source of what we measure.
- Reputation
- What AI associates with your brand right now, per assistant, set against the goal you defined in your profile.
- Visibility and Citations
- How often ChatGPT, Claude, Gemini and Perplexity mention and recommend you on the topics you follow, plus which domains they read for it.
- Retrieval
- What the models look up before they answer: the queries they fire, clustered by subject, and the checks against your own site, with whether those land.
- Conversations
- How a conversation about your category runs and where it tips towards a brand.
- Action Plan
- Everything the analyses produce, stacked by impact and effort. You decide what goes in, you move actions from proposed to done, and you export to CSV or Markdown to hand to a developer or content agency.
- Presentations
- A deck straight from the tool, for when you need to explain it internally or to a client.