When a potential client asks an AI to recommend the best accountant in Turin for a startup, the AI looks for content that genuinely helps them choose — not pages describing what a firm does. If your site explains your services but doesn't answer this kind of question, you never appear for the queries that bring in buyers ready to purchase. You're losing the hottest leads at the exact moment they're deciding. Building the right pages for these search patterns is precise work — and it delivers direct commercial results.
“Recommend the best accountant in Turin for a startup.” This sentence is not a Google search. It’s an instruction. The user knows what they want, knows where they are, knows the context they operate in — and is delegating the selection to the AI. The model won’t return ten results to explore. It will give one answer. Just one. With a name, or with specific criteria for choosing.
If your content isn’t built to match that pattern — precise service, geographic context, specific need — you don’t show up. And it’s not a question of ranking. It’s a question of linguistic architecture: the model looks for sources that answer the instruction to the letter. Not sources that talk generically about the topic.
This is the mechanics of Instruction Following, and it’s the point where most content strategies fail without knowing it.
The mechanism: why the model follows the instruction to the letter
Instruction Following is not a random emergent behavior of language models. It’s the direct result of how they’re trained. Modern research distinguishes three paradigms: standard pretraining, fine-tuning on specific tasks, and prompting. Instruction tuning occupies a space distinct from all three.
As documented by Minaee et al. (2025), “Fig provides a comparison of instruction tuning with pretrain–finetune and prompting” — and the difference is substantial.
Instruction tuning optimizes the model not to complete a text nor to answer a generic prompt, but to interpret and execute composite instructions. The user isn’t searching — they’re asking the model to do something precise.
This changes everything. When a model trained with instruction tuning receives “recommend the best X for Y,” it breaks the instruction down into distinct constraints: the type of action requested (recommendation, not list), the object category (X), the context of application (Y). Then it looks for sources that satisfy all the constraints simultaneously. Content that covers X but not in the context of Y doesn’t pass the filter. Content that lists options instead of selecting “the best” doesn’t match the requested action.
From this it follows that the granularity with which you write your content is directly proportional to the probability that the model will use it as a source for that type of query. This isn’t a speculative deduction: it’s the direct consequence of the training mechanism.
How instructions are evaluated: the role of the reward
There’s an aspect of the process that is rarely explained outside technical papers, and that changes the way you have to think about the structure of your content.
Models are not trained to follow instructions on a fixed sample of commands. The process is dynamic. Ji et al. (2025) document the approach developed by Honovich et al., where “reward-driven evaluations” are used to evaluate automatically generated instructions.
In practice: the system learns to measure how well a response executes the original instruction, and uses this signal to improve future behavior.
The critical point is that this evaluation is calibrated on fidelity to the instruction, not on the abstract quality of the response. Mediocre content that perfectly matches the instruction pattern can beat excellent content that doesn’t match it. The relevant question isn’t “do I have good content on X?” but “do I have content structured to answer how users ask the AI about X?” The difference is the one between a services page and a page that helps you choose.
Content that lists options instead of selecting “the best” doesn’t match the requested action.
The scope of the instruction following problem
Before thinking about the solution, it’s useful to understand the size of the challenge. Beltagy et al. documented it directly in the paper “How far can camels go? Exploring the state of instruction tuning” as a reference point for understanding the limits of instruction following in current models.
Instruction-tuned models perform better on the patterns seen most frequently during training — and high-value commercial queries (“best X for Y”, “compare A and B for Z”, “recommend X in Y”) are among the most recurrent in real conversations. The model was optimized precisely for these cases.
From this it follows that building content around these patterns is not an artificial stretch. You’re aligning the structure of your content with the patterns the model excels at. You’re not writing for the algorithm — you’re writing in a format that the user uses and that the model knows how to handle well.
A page titled “How to choose the best SEO consultant for your ecommerce” must actually provide selection criteria, distinguish between use cases, and give an explicit recommendation for each client profile.
The high-intent queries you’re losing
The queries that follow instruction following patterns are not generic. They are almost always queries with high conversion intent: the user has already defined the problem, has already narrowed the field, is looking for someone who satisfies precise constraints. “What’s the best [service] for [need]?”, “Compare [A] and [B] for [context]”, “Recommend [service] in [city] for [sector]” — each pattern requires a different type of content.
A standard services page doesn’t answer any of these patterns. It describes what you do — it doesn’t help the user choose. This is the structural problem of many sites: they have content that answers the question “what do you do?” but not the question “how do I choose the right one for me?”. The first format doesn’t match the instruction pattern. The second does.
The crux of deduction: what the model does, and what your content must let it do
Here comes an element worth separating out precisely. The AI doesn’t select sources passively — it reasons about them. As I explained in the articles on Chain-of-Thought and Planning, the model builds multi-step reasoning before formulating the final answer.
When it receives a query like “recommend the best consultant for a B2B ecommerce in Milan,” the model deduces: which criteria distinguish a good consultant for B2B from one for B2C? How does geographic location factor in? What makes a choice “better” in this context?
Your content must provide the building blocks for this deduction. It’s not enough to say that you work with B2B ecommerce in Milan. You must provide explicit criteria the model can use to build the recommendation: what makes your approach more suited to that context, which problems you solve that others don’t address, how the result is measured in that specific sector.
Content that contains explicit selection criteria — “choose X if you need Y, consider Z if your main constraint is W” — is constructively more useful to the model than content that simply states “we are experts in B2B ecommerce.” The first feeds the model’s deductive reasoning. The second is a declaration the model doesn’t know how to use.
The connection between instruction following and reasoning becomes even more critical when looking at Hallucination: when the model doesn’t find explicit criteria, it fills the gaps with generalizations. Providing the criteria reduces this risk and increases the precision with which the model represents you.
How to structure content to match the pattern
The principle is simple: every page you want to appear in instruction following queries must be built around a specific instruction pattern — not around a service description.
The first step is to identify the real patterns. Think about the five questions your potential clients would ask ChatGPT when they’re ready to choose — not when they’re still exploring. Not “what is SEO,” but “what’s the best SEO consultant for an ecommerce doing 500k in revenue.” The difference is in the buyer journey stage, and instruction following AI queries are almost always in the selection stage.
The second step is to build pages that answer those patterns, not that cite them. A page titled “How to choose the best SEO consultant for your ecommerce” must actually provide selection criteria, distinguish between use cases, and give an explicit recommendation for each client profile. The pattern must be visible in the structure of the content, not just in the title.
For geographic queries — “recommend X in Y” — the content must contain both the service and the location explicitly and in context, not just as a keyword. If you operate in Milan, mentioning Milan once isn’t enough: the content must explain what changes for a client in that specific market.
For comparative queries — “compare A and B for Z” — the page must have an explicitly comparative structure, with criteria, pros/cons, and a clear recommendation. The model is trained to answer comparison requests with comparison structures: if your content has that form, the model reuses it. On how models handle local verification tools in these workflows, the article on Tool Use explores the specific mechanism in depth.
Practical check: where you are today
Before changing anything, run this check: take the three most likely instruction following queries for your sector and test them on ChatGPT, Perplexity, and Gemini. Don’t ask for generic information — use the exact pattern a potential client of yours, ready to choose, would use.
Do you appear in the answer? How often? In what position? If you don’t appear, analyze the content the model cites: is it structured like a recommendation? Does it contain explicit criteria? Does it have the geographic or sector context integrated into the body of the text?
This tells you where to intervene first. Not across the whole site — on the pages that guard the instruction patterns with the highest commercial value for your business.
The AI follows instructions to the letter. Your content must match those instructions with the same precision. It’s the minimum requirement to appear where it counts.
