The Context Gap: Why Your AI Marketing Sounds Generic

The short answer

A language model writing about your business knows a great deal about businesses in general and almost nothing about yours. Whatever you do not tell it, it fills with the most typical answer. Ask for a post about "why customers choose us" with no evidence, and you get the reasons customers choose anyone: quality, service, expertise, value. That is the context gap, and no amount of prompt wording closes it; only context does.

There is research evidence for the averaging effect. In a 2024 experiment published in Science Advances, Doshi and Hauser found that writers given AI-generated ideas produced stories rated as more creative individually, but those stories were more similar to each other than stories written without AI. Individually better, collectively more alike: that is exactly what generic marketing feels like from the reader's side, when every competitor uses the same tools with the same thin prompts.

The competitor swap test

Take any AI-assisted piece and replace your company name with a competitor's. Read it again. Every sentence that is still true is generic.

A first draft written from a thin prompt often fails on most of its sentences. That is useful, because the sentences that survive the swap point to where your real differentiation is, and the ones that fail show which kind of context was missing.

The seven kinds of context a model cannot guess

Context inventory: what the model does not know

Context	What it is	Where it lives	What it changes in the copy
Customer language	The exact words customers use for their problem and your solution	Sales call notes, reviews, support tickets, enquiry forms	Headlines and openings match what readers search and say
Specific proof	Results, cases, numbers and details you can substantiate	Case records, testimonials with permission, internal reports you may publish	Claims become checkable instead of adjectives
Point of view	What you believe that others in your market do not	Founder interviews, sales objection handling, internal debates	The piece argues something, rather than surveying
What you will not do	Customers, work and tactics you turn down, and why	Proposal notes, lost-deal reasons	Positioning becomes sharp by exclusion
Offer mechanics	How working with you actually happens, step by step	Onboarding documents, process maps	Removes vague promises; answers the how question
The real alternative	What the customer would do if you did not exist	Sales calls, win and loss notes	Comparisons address the actual decision
Recent specifics	What changed this month in your market or business	Team updates, industry news you follow	Content is timely and could not have been written last year

A model has none of this unless you supply it. Google's guidance on people-first content asks whether content provides original information and demonstrates first-hand expertise; every row above is a source of exactly that.

Why better prompts are not enough

Prompt techniques change how the model uses what it has. They cannot create facts about your business. A prompt that says "be specific and original" with no specific, original material gives the model an instruction it can only satisfy by inventing details, which is worse than generic.

Anthropic's own prompting guidance makes a related point: explaining the context and motivation behind a request helps the model target its answer. Context is the input; the prompt is the instruction for using it.

Building a context pack

A context pack is a reusable document you attach to every content request, so context is supplied once and maintained, not retyped. Keep it to what changes the copy.

1. • Customer language: 20 to 30 verbatim phrases from calls, reviews and enquiries, grouped by problem.
2. • Proof library: each claim you can make, with its source and any conditions (timeframe, client permission).
3. • Point of view: five to ten opinions, each in one sentence, with the reasoning.
4. • Exclusions: who you are not for and what you will not claim.
5. • Offer walkthrough: the steps of working with you, in plain words.
6. • Alternatives: what prospects compare you with, and your honest view of when the alternative is better.
7. • This month: three to five recent specifics, refreshed monthly.

Start with the two rows where you are thinnest; often that is customer language and point of view, because they live in people's heads and calls rather than in documents.

Illustrative example

Illustrative example, not a client. A payroll bureau asks for an opening paragraph for a service page.

Without context: "We provide reliable, accurate payroll services tailored to your business, so you can focus on what you do best."

With a context pack (customer phrase: "I dread the last week of the month"; exclusion: no businesses over 250 staff; proof: named accreditation it holds): "If the last week of the month is the one you dread, that is usually because payroll depends on one person and one spreadsheet. We run payroll for businesses with up to 250 staff, and the first month is spent moving that knowledge out of one head and into a process."

The second version passes the swap test on almost every clause. Nothing about the model changed.

How we handle this at AfM

Our service is built on the same principle: the work starts with making context easy for the client to give (calls, notes and existing material rather than long questionnaires), our operating system keeps that context from one piece to the next, we review output against it, and the client approves before anything is published. For the related problem of copy that is specific but sounds mechanical, see why AI content sounds robotic.

FAQ

Will a more advanced model fix generic output on its own?

Only partly. A stronger model writes more fluently, but it still has no access to your customers' words, your proof or your opinions. Without that input, it produces a more polished version of the average answer.

How long does a first context pack take to build?

A usable first version can come from one recorded conversation with the founder or sales lead plus an hour pulling phrases from reviews and emails. It improves every time a reviewer rewrites a draft and records why.

Is it safe to put customer quotes into an AI tool?

Remove names and identifying details first, and check which tools your organisation has approved for business data. Customer language is valuable because of the phrasing, which survives anonymisation.

What if we genuinely are similar to our competitors?

Then specificity has to come from proof, process and point of view rather than from the product. Two firms with the same service still differ in who they serve best, how they work and what they believe.