Chapter 10

One truth, many answers

The map multiplies your messages. The machine reads them all at once. One thing survives both: a story that's true of you and of nobody else, told the same way everywhere.

The arithmetic problem

Go back to the map of buyers and journey steps. Ten kinds of buyer and twenty journey steps make 200 moments. Seven stages in each makes 1,400 answers to get right.

For most of marketing's history that number was absurd. AI has changed the sum. Drafting 1,400 answers now costs little. Checking and approving them still costs what it always did.

Nobody should start with all 1,400. Chapter 12 shows how to choose the few moments worth winning first.

Which leaves a new problem: if you can draft them, so can everyone else.

When everyone can say everything

When any brand can produce any answer, answers stop setting brands apart. The scarce thing becomes what's true only of you: your data, your proof, your method, your history, your point of view.

There's early evidence that machines reward this. A Princeton-led team tested which rewrites made a page more likely to be used in AI-generated answers, across 10,000 queries (Aggarwal and colleagues). They called the practice generative engine optimisation, or GEO.

Adding statistics, quotations and cited sources lifted a page's visibility by up to 40% on their simulated engine. On Perplexity, a live AI search engine, the gain was up to 37%. Stuffing in keywords did worse than changing nothing.

The tests ran on 2023-era systems, and measured being used in an answer, not being bought. Even so, specific and evidenced beats generic, and the most specific evidence of all is the kind only you hold.

It also has to be checkable. When the stakes are high, buyers audit the machine, and only proof they can verify survives.

The living answer

How does one story become 1,400 answers without becoming 1,400 stories? Every cell in your body carries the same DNA and expresses it differently. The identity never changes, and the expression always does.

What stays and what changes

Personalisation that changes the right-hand column is good marketing. Personalisation that changes the left gives you a different brand every time.

Marketers have called this discipline integrated marketing communications since the 1990s. What's new is the number of answers, and a reader that sees them all.

One truth, held in one place

Saying one thing in 1,400 forms can't be done from a brand deck. It needs a single structured account of who you're for, what you claim and the proof, which every answer is drawn from.

The evidence for structure comes from inside a company's own systems. In one benchmark, GPT-4 answered insurance-company database questions correctly about 17% of the time. Once the same data was organised as a graph of facts and relationships, it was right 54% of the time (Sequeda and colleagues).

The test was internal business data, not brand content. I read it as a reason to keep your own facts structured, not as evidence that public AI assistants read your structured data. Google says no special markup is needed for its AI answers. In one test, assistants treated a site's hidden structured data as ordinary text (Search Engine Journal).

So the one truth is the foundation, not the answer. Its job is to make everything you publish agree, so the machine meets one consistent brand, in words it can read and others can confirm.

Fit the question, keep the facts

The right-hand column matters too. Researchers at Columbia and Yale built a simulated marketplace and let AI shopping agents choose among products (Allouah and colleagues).

A seller's own agent then made "simple, query-conditional description tweaks": it reworded the product description to fit the question being asked. The facts stayed fixed. Across six AI models, the product's average share of the agents' choices rose by between 0.32 and 14.89 percentage points.

The gains were uneven: a single round of rewording produced a significant increase in only 33% of the experiments. Fit to the question helps, by an amount that depends on which model is choosing.

It was a simulation. But it shows both halves of this chapter at once. Change the words to fit the moment. Never change the facts.

Why the machine raises the stakes

A human buyer only ever saw their own moment. The machine reads many at once, but it's a poor referee. Given two sources that disagree, AI models mostly fail to notice the conflict, and pick one version to state with confidence (Chapter 16).

Tell one audience your software saves 20% and another 30%, and the machine may repeat either, to anyone who asks. The buyer who checks and finds the other version now doubts both.

Nothing yet shows that consistency on its own makes a machine recommend you more. The case rests on what inconsistency costs: a lapse can now be repeated at scale.

This afternoon, this quarter

This afternoon: write one page. Who you're for, the three things you claim, and the proof for each. Then hold your home page, your best-selling product page and your last proposal against it, and mark every place they disagree.

This quarter: turn that page into a structured record that your team and your tools draw every answer from. Make it somebody's job to keep it true.

The next chapter gives that truth an edge.

Whose evidence this is. Sequeda and colleagues work for data.world, which sells data-catalogue and knowledge-graph software. Search Engine Journal is a trade publication. Academic sources: Aggarwal and colleagues (presented at KDD 2024, a peer-reviewed conference); Allouah and colleagues (a preprint, run as a simulation).