The Inclusion Problem

Same need, different words, different answer

In September I tracked the questions UK banks ask when they're choosing a partner to modernise their payments systems, across ChatGPT, Gemini and Google AI Overviews. Two of them ask for the same thing: "Who can help meet the ISO 20022 deadline?" and "Best partner for ISO 20022 migration".

One of the world's largest IT services firms was named in 67% of answers to the first question. In answers to the second, it wasn't named once.

Global IT services - same firm, different question

Nothing about the firm changed between those two questions. Only the buyer's wording did. That's the Inclusion Problem in one example: before AI can recommend you, it has to put you in the answer, and it makes that call separately for every question it's asked.

Where does Inclusion sit in the buying journey?

The Selection Layer is where AI shortlists get built. An AI shortlist is the handful of vendors an AI tool recommends when a buyer asks who they should consider.

The Selection Layer has four layers - Inclusion, Positioning, Preference and Selection - and Inclusion comes first because everything else depends on it. A brand that isn't in the answer isn't framed, prioritised or chosen. It simply isn't there.

In How B2B Buyers Actually Buy in 2026 I set out the five stages of the journey. AI does most of its work in the third, the Comparison Moment: list in, shortlist out. Some buyers arrive with a Day One List and use AI to cut it down. Others arrive cold and let AI tell them who the players are. The shift is happening fast: the share of software buyers starting their research in an AI chatbot rather than Google nearly doubled in a year, from 29% to over half (G2, 2026).

Whichever route they take, the buyer still makes the decision, but they make it from the shortlist the AI puts in front of them. And AI can only shortlist the brands it has already put in its answer.

So how does AI decide what goes into its answer? These tools are trying to give the best answer to the question in front of them. The best answer is the one with the highest probability of satisfying what the model reads as the buyer's requirements and intent, based on how they asked. To build it, the model looks across a wide range of sources for building blocks it can put together, and it favours the sources it believes it can trust. AI assembles the market from what's already been written about it.

Whether your brand gets named in what it assembles comes down to three boxes, and you have to tick all three to be included.

What decides whether a brand is included?

Recognition: does AI know you exist as a player?

I'll start with us. When a question names Decoding Growth, AI includes us about two-thirds of the time. When it doesn't, we turn up in roughly one answer in 200, and Google's AI Overviews haven't named us once.

Decoding Growth is a new name, so failing Recognition first is no surprise. It still means a prospective client who asks AI for help with exactly what we do would almost never hear our name, which is why we track it, and why we're working to fix it.

Retrievability: can AI find and read what you do?

One of the big card networks runs a payouts service for businesses. AI names it in 84% of answers when buyers ask about card payouts. When they ask about cross-border, marketplace and treasury payouts, that drops to 14%. In those cross-border answers, the models drew on the network's own website 4% of the time. Explainer pages from Freysoft, a small fintech and payments software engineering firm, were used four times as often, in 16% of answers.

Size doesn't decide whether AI can use your content. Pages that directly answer the question the buyer asked do. If the answer to the specific question lives on someone else's site, that's where the model builds from.

I saw the same thing last week with an enterprise architecture software vendor. Ask AI for a platform in their category, or for an alternative to their biggest competitor, and they're in every answer, and first whenever the buyer names that competitor. Describe the problem their product solves instead, such as meeting the FCA's operational resilience rules, and most engines don't send the buyer to enterprise architecture tools at all. In the four answers that did, the competitor was named every time and the vendor twice. Twice, an engine leaned heavily on a competitor's own pages about the buyer's exact problem. In neither case was one of the vendor's own pages cited.

Corroboration: does anyone else vouch for you?

Credibility and authority aren't determined by what you say - they're determined by what others say about you.

I learned how this works at Spiceworks, long before AI was part of the picture. It was then the world's largest community for IT professionals, and its value came from the questions IT pros asked each other about the products they were about to buy: what did you choose and why, what's working, what should I watch for, what went wrong. They didn't trust what vendors said on their own sites, because it was inherently biased. They wanted an impartial view from people they felt had more credibility and authority.

In How B2B Buyers Actually Buy that's the Trust Loop: the buyer checking AI's answer with their peers after they've read it. What's changed is that AI now runs the same check before it writes the answer. It does exactly what those IT pros did - at speed.

When I tracked the UK business laptop market in July, the brands’ own websites barely featured. The answers were built from TechRadar, RTINGS, PCMag and YouTube, and one tech-media article was cited more often than the four brand websites I reported on put together. Laptops are a review-media category, though. The B2B software and services categories I’ve tracked since behave differently: vendor websites make up much of what the AI reads there - including your competitors’.

Buyers weigh evidence the same way. 85% trust peer recommendations and 78% trust third-party reviews, but only 32% trust a vendor's own thought leadership (Omniscient Digital/Wynter, B2B SaaS buyers).

The enterprise security data showed something sharper: being read isn't the same as being named. One vendor's own website fed dozens of AI answers about how to solve a problem, and the vendor was named in about one in six of them. Gemini answered a question on securing SaaS applications from another vendor's glossary, cited it five times, then recommended two competitors.

The obvious response is to publish more. These vendors already had: their content was being read and used, and it still didn't get their name into the answer.

This isn't a quirk of our sample. A Semrush and Growth Memo study found that in six out of ten cases where a site appeared in an AI answer, it was cited without the brand behind it being mentioned. Explainer content gets cited; comparison content is far more likely to get you named.

Of the three conditions, this is the hardest. You can build Recognition and fix Retrievability on your own site. Corroboration you have to earn somewhere else - and in B2B, “somewhere else” is often another vendor’s site. In my October HR tracking, a holiday planner was recommended in 12 answers without its own site being cited once. Three of those answers linked its name to a rival vendor’s “top 8” list (see the note on method).

Why can't you see it from where you sit?

There's no Search Console for AI answers. You can't see what buyers type into ChatGPT, and neither can anyone else. Every tracking tool, including Peec, which I use, runs its own prompts and samples what comes back. That gives you a repeatable baseline, not a census, and personalisation moves individual answers around it.

The clicks won't tell you either. Pew Research found only about one in a hundred visits with a Google AI summary led to a click on a link inside it. Buyers take the answer and act on it somewhere else, so most of the influence never shows up in your analytics.

And the engines disagree. Growth Memo found that nine in ten cited pages turn up on only one of ChatGPT, Perplexity and Google AI Overviews for the same question. Our own numbers differ by engine too. One engine is one opinion.

What should you do this week?

Run the test yourself, and make it as close to a real buyer as you can.

1. Write down five questions your buyers actually ask, in their words - including the specific ones with a constraint, an integration or a deadline in them, and at least one pair that asks the same thing two ways.
2. Open a private window and log out of every AI tool, so your history and memory don't colour the answer.
3. Run each question on ChatGPT, Gemini, Perplexity and Google's AI Mode, and note who's named and which pages are cited.
4. Run them again next week. AI tools won't give you the same answer twice, so one run tells you very little.
   

Check what you find against the three "How you'd spot it" signs above. It won't give you a score, but it will show you which condition is holding you back.

If you'd rather have it done properly - thirty buyer questions across five platforms, repeat runs, scored across all four layers - that's what the AI Shortlist Scan does.

Run the specific question this week, before your next buyer does.

Next steps

Want to know where your brand stands? The AI Shortlist Scan is a decision-grade analysis of how AI is currently representing your brand across all four layers of The Selection Layer. £1,995, delivered in seven days, money-back guarantee.