Intent-based content: how to create content for buying intent in AI search

Your customers are less and less likely to type in keywords, and more likely to describe their situation. “What warehouse management system should a company with 20 employees and a budget of up to €7,000 use?” isn't a keyword. It's a prompt, one the model will answer with a recommendation, not a list of links. If your content doesn't answer that exact description of the situation, it won't appear in that recommendation, even if you rank first in Google.

This piece shows you how to shift from thinking in keywords to thinking in buying intent when creating content, and exactly what to do with your existing articles so AI models can pull answers out of them.

What Is Intent-Based Content in AI Search?

Intent-based content is content built around the intent behind a query, not around the query itself. In an AI context, this means one more thing: content has to serve individual sub-intents at paragraph level, because generative models assemble their answer from fragments of many sources, not from a single best-matching page.

The difference is practical. In classic SEO, it was enough for the whole page to match the intent of a keyword. In AI search, what counts is whether a specific piece of text answers one specific sub-question that the model generated for itself.

Search Intent vs. Intent in AI: What Exactly Has Changed

Three things have changed: the length of the query, how it's processed, and the unit the algorithm evaluates.

Semrush data from an analysis of more than 80 million clickstream records shows the scale of this difference: only 30% of ChatGPT prompts fit the classic intent categories known from search engines. The remaining 70% are queries that simply don't show up in classic keyword research.

Why Intent Stopped Being a Property of the Keyword and Became a Property of the Paragraph

Because the model doesn't pick a page, it picks fragments, and it does that separately for each sub-intent. One of your sections might get used for a question about selection criteria, another for a question about costs, and a third might not be touched at all, even though it sits in the same piece of text.

Data from llmpulse.ai from April 2026 shows that 44% of citations come from the first 30% of a page's content. Data and figures tucked away in the second half of an article are practically invisible to the model. This flips the classic logic of writing, where conclusions and specifics used to go at the end.

The consequence for your content work: you stop optimising articles and start optimising sections. Each one is a separate unit that has to make sense without the rest of the article.

Intent-Based Content, GEO, AEO and SEO: How They Relate to Each Other

These aren't competing approaches, just different layers of the same work.

• SEO answers the question of whether a page can even be shown at all, and whether it's technically accessible to crawlers.
• GEO and AEO answer the question of whether content is fit to be cited in an answer generated by a model.
• Intent-based content answers the question of whether you're being cited on queries that have commercial value, rather than random ones.

You can be cited and not earn a single penny from it. That's why the intent layer is the one that determines the return on everything else. Semrush data from its 2026 AI visibility index also shows that companies running SEO and AI visibility as one joined-up strategy saw 81% growth in visibility, versus just 36% for those treating them separately.

Why Have Buying Intents Moved to AI Models?

Because models have stepped into exactly the moment in the journey where you decide which supplier to choose. Not the moment of purchase, but the moment of comparing and narrowing down options, which is exactly where comparison content and guides used to do their job.

The scale here is already measurable, not just forecast. In February 2026, ChatGPT had about 900 million weekly active users, twice as many as a year earlier. According to the Adyen Index: Retail Report 2025, 37% of Poles already use AI in the purchasing process, and 49% are ready to.

How Often Do AI Overviews Appear for Commercial Queries

This is growing fast, and unevenly across industries. Semrush analysed more than 600,000 keywords across 10 industries between November 2025 and April 2026, and recorded a 71% increase in the presence of AI Overviews on results with commercial intent.

The value of these queries is changing too. In finance, the average CPC for keywords with an AI Overview was $4.84, against $2.14 for keywords without one. In education and jobs, it was $5.02 and $1.51 respectively. Google doesn't put AI Overviews on queries with no value: it puts them where buyers hesitate the longest.

How Does Traffic From AI Convert Compared With Organic Traffic

Better, but how much better depends on who you ask. To be fair, we need to give a range, because individual studies differ by an order of magnitude.

The most interesting thing in this table is the trend, not the value itself. In March 2025, according to Adobe, traffic from AI converted 38% worse than other channels. In March 2026, it converted 42% better. That's a shift of around 80 percentage points in twelve months, and the strongest available argument that we're looking at a change in behaviour, not statistical noise.

The behavioural profile fits too. According to Adobe, users coming from AI assistants spend 48% more time on the page and visit 13% more subpages. They arrive on the site after the research stage, to confirm price, specifications and credibility. These aren't people starting their search. They're people finishing it.

Why Transactional Queries Behave the Opposite Way to Commercial Ones

Because AI helps you choose, not buy. In the same Semrush study, where the presence of AI Overviews on commercial queries rose by 71%, on transactional queries it fell by 5%.

That's an important nuance, and it changes your content priorities. When someone types a product name with intent to buy, Google doesn't insert an AI summary, because a product page and shopping results are enough. The summary appears earlier, on questions like “which one should I pick”, “how do they differ” or “is it worth it at this scale”.

The conclusion is simple: the fight for visibility in AI is happening at the choice stage, not the checkout stage. Content that serves this stage, comparisons, criteria, limitations and use cases, has higher value today than yet another article targeting a purchase-intent keyword.

Three Numbers Circulating Online That Aren't Worth Repeating

The AI search market is flooded with figures that fall apart under scrutiny. The three most common:

• “Search volume will drop by 25% by 2026”. This is a Gartner forecast from February 2024, not a measured result. It hasn't materialised at that scale, yet it's still being cited as fact in articles.
• “Traffic from AI converts 23 times better”. This figure comes from one company's own sign-up data for a SaaS product. It's true for that one case and useless as a benchmark. The real market range is anywhere from 1.2x to a few dozen percent better conversion, depending on the industry.
• “90% of ChatGPT citations come from pages ranked 21 and below”. Senuto data from 18 million keywords shows the opposite: the average rank of a cited source is 10.06, the top 3 accounts for 16.78% of citations, and the top 10 for 49.28%. The correlation between Google ranking and being cited in AI is about r = 0.65.

The third figure is the most damaging, because it leads to the conclusion “ditch SEO, do GEO”. The data says a good organic ranking is a prerequisite, not a relic.

How Does an AI Model Work Out User Intent?

The model breaks the query down into parts, searches separately for each one, and only then assembles the answer. This mechanism is called query fan-out, and it's the single most important thing to understand if you want to write for AI deliberately.

The process has four stages: decomposing the query into entities, constraints and time references; parallel searches for each sub-query; expanding to further sources; and finally refining the intent and synthesising a single answer.

Query Fan-Out: How One Question Turns Into 8 to 12 Queries

For a standard query, Google AI Mode generates 8 to 12 sub-queries, and in deep search mode, even hundreds. The number depends on complexity: the model checks how many named entities there are, whether a comparison shows up, and whether there are clear constraints.

Take the question “which ERP system suits a manufacturing company with 50 employees and a budget of up to €25,000?”. The model breaks it down into questions about typical ERP features for manufacturing, systems available in that price range, implementation time, hidden costs, user reviews, and requirements for a company of that size.

Your content takes part in this process as many times as it answers these questions. If it answers one, it gets one chance. If it answers six, it gets six. That's why an article that only answers the question in its title is almost invisible in the fan-out, even with a high Google ranking.

Why the Model Pulls Fragments, Not Pages

Because the answer has to be concise, and a page isn't. The engine pulls fragments, evaluates them against a single sub-question, and combines them with fragments from other sources.

The practical result: a fragment that starts with “as mentioned above” or “in this context it's worth adding” is useless. Pulled out of the text, it means nothing, so it won't be used.

The data shows just how big the premium for good structure is. According to Mention Network, fragments with the answer up front collect about 3.8 times more citations. Sections that are 120 to 180 words long get about 70% more than sections that are clearly longer or shorter. Tables are cited 2.5 times more often than the same material presented as running text.

Conversation as Accumulated Intent: The Model Asks About What the User Didn't Say

A conversation thread acts like a memory of intent, so each subsequent question gets narrower and more purchase-focused. The user doesn't have to repeat the context: the model knows it's about a manufacturing company with a set budget, and narrows the recommendations itself.

This is where the difference emerges between content that wins the first question and content that wins the whole conversation. The first question is usually broad, so the answer comes from general sources. Later ones already deal with variants, exceptions, limitations and costs. Whoever has these described explicitly wins them.

That's why content built for buying intent has to include things marketing usually avoids: who the product isn't suited for, where its limits of use lie, what drives up implementation cost, and what the standard offer doesn't cover.

What This Means for Text Structure: Every Section Has to Work on Its Own

The test is simple: copy any section into a blank document and check whether you can pull a sensible answer out of it without the rest of the article. If you can't, that section won't work in AI.

In practice, this means four habits. A heading phrased as a question or an unambiguous topic. The answer in the first sentence under the heading. Proper names and numbers stated explicitly in the text, not implied. No references back to earlier paragraphs.

What Do Real Purchase-Intent Prompts Actually Look Like?

Shorter and less flashy than prompting guides suggest. A Stella Rising study from January 2026, based on a sample of 524 active users of language models with a margin of error of ±4.3%, shows the median prompt length is 12 words, and 66% of queries fit within 15 words.

This matters, because a large share of content being created today “for AI” is optimised for imagined, elaborate prompts with a persona and detailed instructions. The real user writes more simply.

A Median of 12 Words, 60% Phrased as a Question

About 60% of queries take the form of a question, and only 9% take the form of a direct command. That's good news, because a question is easier to predict and to mirror in a heading.

There's also a difference depending on whether the model uses search. According to Semrush data, prompts without search mode enabled average 23 words and have an exploratory character, while those with search average 4.2 words and resemble classic queries. In practice, this means short keywords haven't disappeared, descriptive queries have simply joined them. You need to serve both.

More than 90% of monitored prompts trigger a live fetch of content from the web. So visibility in answers isn't purely a matter of what the model “remembers” from training. You can influence it right now.

Personal Context in 32% of Prompts: Age, Occupation, Situation

In almost a third of queries, the user shares something about themselves, not about the product: age, occupation, health, life situation, company size, stage of business growth.

This is the biggest difference from keyword research, and the biggest opportunity. The keyword “work shoes” says nothing. The prompt “what shoes should I wear for a job where I stand for 10 hours and I have back problems” says everything, and practically tells you what content will win.

Other common elements of prompts in the same study: a budget constraint or price in 28% of queries, the word “best” or an equivalent in 24.5%, and local context in 16%.

The conclusion: content built for buying intent has to describe the reader's situation, not just the product's features. A “who this is for, and who it isn't for” section matters more today than a “why choose us” section.

How a Purchase-Intent Prompt Differs From a Keyword

A keyword is a label for a topic; a prompt is a description of a situation with constraints.

The practical consequence: keyword research remains the foundation, but it stops being enough on its own. You need a second list, made up of real situations in which someone reaches for your product.

How to Build Your Own List of Purchase-Intent Prompts for Your Business

Start with the sources you already have, not with tools. The best prompts are sitting in your archive of customer conversations, not in keyword databases.

An order that works:

• Sales conversations and quote requests. Write down the questions customers ask before deciding, word for word in their own phrasing. These are ready-made prompts.
• Search Console. Filter for queries starting with “how”, “is”, “which”, “what's better” and “how much”. These are question-form queries, and their share is growing.
• People Also Ask and Google's “More questions” section. Real user questions, especially useful for FAQs.
• Queries from chat, emails and comments. This is where you'll find the objections nobody types into Google.
• Forums and social platforms. Worth checking, because Reddit accounts for 46.7% of citations in Perplexity, meaning the model is literally reading these discussions.
• Prompt and AI visibility monitoring tools. Only at the end, to verify and measure, not to come up with the list.

Add one thing to each prompt: what stage of the decision the person asking it is at. This lets you map content to a stage, which is what the next section covers.

A Map of Buying Intent in AI: Five Stages and the Content That Serves Them

The buying journey in AI can be broken down into five stages, each generating a different type of prompt and requiring a different type of content, placed in a different location. Below is a map of the whole thing, followed by a breakdown of each stage.

Stage 1: Problem Recognition, When the User Doesn't Know the Solution Yet

At this stage, the user describes a symptom, not a product, so the content that wins is the one that names the problem in their own words. The prompt is “why are my storage costs so high”, not “WMS system”.

The model is looking for causes, mechanisms and ways to diagnose the issue here. It reaches for explanatory content and external sources, because it doesn't trust sales material in this role. According to xfunnel data, at the top of the funnel the share of product and vendor content falls, while the share of research and industry publications rises.

Your job: describe the problem without selling the solution in the first paragraph. The section that collects the most citations at this stage is a list of causes, each with a short explanation and a way for the reader to check which one applies to them.

The signal it's working: an increase in brand mentions on broad problem-focused queries, even before any traffic shows up.

Stage 2: Building Selection Criteria

Here, the user already knows what they're looking for, but not what to base their choice on, so they ask about specs and pitfalls. This is the most underrated stage, because it's where the rules get set that the model will later use to compare offers.

Whoever supplies the criteria sets the frame for the comparison. If, in your industry, what really matters is implementation time and running cost, but everyone writes about features, then describing those two criteria explicitly, with numbers, tilts the comparison in your favour.

The format that works here: a specification table showing what each value means in practice, plus a clear note on which criterion matters at which scale of business. Models happily reach for content with a high density of facts. According to a Surfer SEO analysis, cited fragments had a median of 0.41 facts, versus 0.15 for uncited ones.

Stage 3: Comparing Options and Alternatives

This is the stage with the highest commercial value, and the hardest one, because the model barely trusts a single source here. Prompts come in three recurring forms: “X or Y”, “best X for a specific use case”, “alternatives to X”.

For broad comparison queries, about 85% of citations come from third-party sources, not the manufacturer's own domain. According to the same data, domains with profiles on review aggregators have about 3 times higher odds of being cited in ChatGPT.

This doesn't mean your own comparison is pointless. It means it has to be written differently from sales material: with explicit criteria, with the real drawbacks of every option, including your own, and with a clear statement of who's better served by the competing option. A piece of text that wins every single comparison gets dismissed as untrustworthy.

The signal it's working: the brand starts showing up in answers to comparison prompts where it didn't appear before.

Stage 4: Validation, Risk and Objections

At this stage, the user is looking for a reason to back out, and the model helps them do that. Prompts cover drawbacks, hidden costs, reviews, implementation risk, and what happens if the choice turns out to be wrong.

Content that serves this stage is an extensive FAQ, total cost data, terms and exceptions. Freshness has a measurable effect here: according to Rampiq data, up-to-date reviews translate into about 28% more citations.

The most common mistake is companies stripping out anything that sounds negative from their content. The model then fills the gap from other sources, usually less favourable ones. It's better to describe the limitations yourself, in your own words and with context.

The signal it's working: fewer recurring objections in sales conversations, and longer on-site sessions from AI traffic.

Stage 5: The Moment of Decision and Capturing Traffic From AI

At the bottom of the funnel, you're back on home turf. According to xfunnel data, more than 70% of citations on bottom-of-funnel queries come from product and offer content, meaning your own site.

The model needs specifics here: a price or a range, availability, scope, delivery time, terms. If your content doesn't have this, the recommendation goes to whoever states it explicitly. It's worth remembering that data infrastructure plays a role here too, not just text: for products, models draw on structured catalogue data, and when comparing offers for the same product, they look at availability, price and how complete the description is.

The purchase itself is a separate matter. Instant Checkout in ChatGPT, launched in September 2025, was withdrawn in March 2026, because fewer than 15 Shopify stores used it. The model shifted back towards product discovery and handing the buyer off to the store. For content, this means one thing: the role of the landing page hasn't disappeared, quite the opposite, it's still the one that closes the sale.

Who Gets Cited at Which Stage: Your Site or Third-Party Sources?

The lower down the funnel, the bigger the role your own domain plays. The higher up, the more it matters whether other people are talking about you. This is the single most practical conclusion from the available citation research, and it's what should decide where your content budget goes.

Data: xfunnel, analysis of citations in ChatGPT, Gemini and Perplexity.

Bottom of the Funnel: Why Your Own Product Page Wins

Because at this stage, nobody but you has the data that's needed. Price, availability, variants, implementation time and terms are information the model won't find in an industry article.

This gives you a very concrete priority: a service page or product page is part of your content strategy, not just a store template. It needs full sentences with facts, not just specs in an image, because the model can't read the latter.

It's also worth checking that the page doesn't open with education. A user arriving from a model is already informed. An intro explaining what the product category is only pushes them further away from the information they came for.

Top and Middle of the Funnel: Without Citations Beyond Your Own Domain, You Don't Exist

For broad queries, models deliberately gather multiple independent voices, so your own content alone isn't enough. About 85% of citations on such queries come from third-party sources, and Reddit accounts for 46.7% of citations in Perplexity.

There's also a difference between markets: for business queries, models more often reach for the manufacturer's own sources; for consumer queries, they mix them with third-party opinions.

In practice, this means three actions in parallel: keeping your profiles current on review aggregators and industry directories, building expert presence in publications that models cite, and keeping your company data consistent across the web. Conflicting information weakens the brand's credibility as an entity.

How to Split Your Content Budget Between Your Own Content and Presence Elsewhere

Start at the bottom of the funnel, because that's where the return is fastest and most measurable. Only then invest further up.

An order I consider sensible on a limited budget:

• Tidying up product and service pages so they state facts, prices and terms explicitly.
• One strong comparison or roundup of alternatives for your most important category.
• An extensive FAQ that handles real objections from sales conversations.
• Profiles on review aggregators, and work on keeping reviews fresh.
• Top-of-funnel content and presence in external publications.

Doing it the other way round, starting with educational articles, is the most common mistake. It builds visibility on queries that third-party sources win anyway, and it doesn't close any sales.

What You Can't Make Up for With Content: Fresh Reviews and Trust Signals

Some of your visibility in AI doesn't depend on text, and that needs saying plainly. Models look at how fresh your reviews are, how consistent your company data is, and whether the brand is mentioned in sources considered credible.

Semrush data also shows a limit on visibility itself: on Gemini, up to 30% of brand mentions have no accompanying link citation. The brand gets recommended, but there's no traffic from it. ChatGPT cites an average of 15 sources per answer, Gemini only 3, so your odds of making it into the answer can vary by up to a factor of five depending on the platform.

This needs to be factored into your expectations. Visibility in AI is the first effect, traffic is the second, and not every instance of visibility can be turned into traffic.

How Do You Write a Section That AI Will Pull Out as an Answer?

A section ready to be cited has five features: the answer up front, the right length, self-sufficiency, a high density of facts, and a structure that's easy to break down into elements. Below is each one, with the number behind it.

The Answer at the Start of the Section, Not in the Summary

The first sentence under the heading has to contain the whole answer; treat the rest of the section as elaboration. According to Mention Network, fragments built this way collect about 3.8 times more citations.

This works at the level of the whole piece too. Since 44% of citations come from the first 30% of the content, the most important data, definitions and figures need to sit at the start of the article, not in the summary.

The simplest test: remove the first sentence of each section and check whether the text still answers the heading. If it does, the first sentence was decoration, not an answer.

Optimal Fragment Length and the Self-Sufficiency Test

According to llmpulse.ai, sections that are 120 to 180 words long collect about 70% more citations than fragments that are clearly longer. This doesn't mean short articles, just longer articles cut into sensible, self-contained parts.

The self-sufficiency test takes thirty seconds. You copy the section into a blank document and ask whether you can use it to answer someone who hasn't read the rest. If the section contains phrases like “as I mentioned”, “in the previous paragraph” or “this problem”, you need to replace them with an explicit name.

Numbers, Entities and Proper Names Instead of Adjectives

A model can't cite an adjective. “Very efficient” means nothing; “a throughput of 1,200 units per hour” means everything.

The same rule applies to names. Specific tools, standards, norms, people's names and places build the network of associations models use to recognise what a text is about and for which queries to use it. Generalities don't build that network.

The practical minimum per section: one number, one proper name, one limitation or exception. If a section is missing all three, it probably won't get used.

Tables, Lists and Structured Data: What Actually Increases Citability

According to norg.ai data from March 2026, tables are cited 2.5 times more often than the same material presented as running text. That's why every comparison, specification roundup and overview of options should take the form of a table, not a paragraph.

Structured data works as support, not a replacement for content. The most useful types are FAQPage for question sections, Product together with Offer and Review for products, and Article for expert content.

An important technical condition that's easy to forget: content hidden in tabs, script-loaded accordions, PDF files or images doesn't exist for the model. Technical specifications given only in an image are invisible.

Information Gain: The One Advantage a Model Can't Generate Without You

Anything that can be inferred from what's already on the internet, the model will generate itself, and it won't need to cite anyone. You earn a citation with information that can't be produced without access to the underlying data.

In practice, information gain means your own data from projects and implementations, results with numbers and context, observations from audits, named cases, prices and real delivery times, and descriptions of what didn't work.

It's also the only part of your content strategy that competitors can't copy, because they don't have this data. The good news is that every company has more of it than it thinks. It's usually sitting in reports, proposals and archived conversations; it just never got published.

How Do You Rewrite an Existing Article for Buying Intent?

You don't need to write everything from scratch. In most cases, it's enough to reorder things, add the missing sub-intents, and cut content that doesn't answer any real question.

The process has four steps, and you can get through it on a single article in a few hours.

Audit: Which Sub-Intents Does the Text Skip

Write out the sub-questions a model would generate around your article's topic, and mark the ones the text doesn't answer. That's the whole diagnosis.

The quickest way: ask the article's main title question in ChatGPT or AI Mode, and see what the answer covers. The elements the model considered important that your text doesn't have are your list of gaps. It's worth adding questions from People Also Ask and Search Console to it too.

Typical gaps you see in expert articles: no selection criteria, no cost information, no statement of who the solution isn't suited for, no numbers.

What to Cut

Throw out everything that pushes the answer further away from the heading: intros that build suspense, repetitions of the same argument, and education aimed at someone at a different stage of the decision.

Prime candidates for removal:

• paragraphs starting with “nowadays” and “more and more companies”
• general definitions in a piece written for people who already know the topic
• judgement calls with no data behind them, along the lines of “this is the best solution on the market”
• conclusions repeated in the summary, if the same thing isn't at the start

What to Add

Add what the model can't build on its own: criteria, limitations, variants, numbers and explicit drawbacks. That's usually two to four new sections.

In order of effectiveness: a comparison table or criteria roundup, a section on who the solution isn't meant for, cost or price-range information, an extensive FAQ at the end, and one of your own figures or observations from projects.

If your choice is between adding another 500 words or one table and an FAQ, choose the latter. Volume doesn't increase citability; structure does.

Example: The Same Paragraph Before and After Rewriting

Below is a passage about choosing a solution for a manufacturing company. It's an illustrative example, but it reflects a real pattern seen in most industry articles.

Before:

Nowadays, digitalising production is no longer a fad, it's a necessity. More and more companies see the benefits of modern solutions that let them optimise processes and increase efficiency. It's worth noting that choosing the right system is an individual matter that depends on many factors.

After:

For a manufacturing company with up to 50 employees, three criteria decide it: implementation time, annual running cost, and whether the system handles production orders without extra modules. Licence price is usually less important, since it makes up 20 to 30% of the total cost in the first year. Solutions that need a dedicated implementation rarely make sense at this scale, because the setup cost outweighs the savings from the first two years.

Three sentences instead of three sentences, but the second version answers four sub-questions, gives a number, names the criteria and states outright when not to do something. The first one answers none of them.

How Do You Measure Whether Content Built for Buying Intent Is Working?

Classic ranking reports aren't enough, because part of the effect never goes through a click at all. You need to split measurement into three layers: visibility in answers, traffic attributed to models, and how that traffic behaves on the site.

What You Can't Measure With Google Rankings

You can't use them to measure either presence in AI answers or brand mentions without a citation. Queries with an AI Overview have about 83% zero-click, and in AI Mode that figure reaches 93%. Visibility without a click is the norm, not the exception.

Search Console also doesn't separate clicks from AI Overviews from the rest of the organic results, so you can't get a clean picture out of it. That's a limitation of the data, not an absence of effect.

That's why the first layer of measurement is prompt monitoring: checking whether the brand shows up in answers to a set of chosen queries, broken down by platform. AI visibility tracking tools exist for this, but you can make a start manually with a list of your 20 to 30 most important prompts.

How to Isolate Traffic From AI Models in GA4

GA4 doesn't have a channel for AI by default, so you have to define one yourself. Traffic from assistants usually arrives as a referral from their domains, so you can catch it with a rule based on session source.

A practical approach: create your own channel group or segment covering sources such as chatgpt.com, perplexity.ai, gemini.google.com, copilot.microsoft.com and claude.ai, and treat it as a separate channel in your reports. It's worth expanding it over time, because the list of tools keeps changing.

Three caveats to keep in mind when reporting: clicks from AI Overviews will land under organic and won't be separated out, some referrals lose their source, and this channel's volume will be low at first. According to WebFX data, traffic from AI is still a fraction of a percent of the total, with very high growth. So judging it by session count alone is misleading.

Which Events Signal Genuine Buying Intent

Instead of session count, look at bottom-of-funnel events, because that's where this traffic shows its value: a contact form, a quote request, a phone number click, an add-to-cart, a spec sheet download, a visit to the pricing page.

It also makes sense to compare channel behaviour, not just conversions. According to Adobe, users from AI assistants spend 48% more time on the page and visit 13% more subpages than users from other sources. If your AI traffic behaves the opposite way, that's a signal your landing page doesn't match the stage these people are coming from.

How Long Until You See an Effect, and What Counts as Normal Variance

The first changes in visibility in AI answers show up faster than in classic SEO, because more than 90% of prompts trigger a live content fetch. Changes in citation can become visible in weeks, not months.

That said, you need to expect a lot of variability. The same answer to the same question can differ between sessions and between platforms, and ChatGPT cites an average of 15 sources, Gemini 3. A single measurement means nothing; what matters is the trend across a repeatable list of prompts.

A realistic expectation: your first citations within a few weeks of publishing or rewriting the content, and stable presence after several months of consistent work on trust signals.

The Most Common Mistakes When Creating Content for Buying Intent

• Optimising for imagined prompts. The real median is 12 words, not a paragraph-long instruction. Instead, write for queries in the form of short questions with one constraint.
• Treating GEO as a replacement for SEO. The average rank of a cited source is 10.06, and the top 10 accounts for close to half of all citations. Instead, run one combined strategy: companies that merge both areas saw 81% growth in visibility, versus 36% for those keeping them separate.
• Stripping drawbacks and limitations out of your content. The model will fill them in from less favourable sources. Instead, describe them yourself, with context.
• Putting conclusions at the end of the text. 44% of citations come from the first 30% of the content. Instead, put the answer at the start of the section and the most important data at the start of the article.
• Adding volume instead of structure. Instead of 500 extra words, add a table and an FAQ.
• Data only in images and PDFs. The model can't read that. Instead, repeat the specifications in text or in a table.
• Starting with top-of-funnel content. Third-party sources win there anyway. Instead, start with product and service pages, where more than 70% of citations come from your own domain.
• Judging the effect by session count. Traffic from AI is a fraction of the total, with high growth and high conversion. Instead, look at events and revenue per visit.

If you want to check which queries your brand shows up in AI answers for today, and what's missing from your content to get into recommendations at the choice stage, take a look at our AI Search audit and AI SEO services.

FAQ: Intent-Based Content and Writing for AI

Also Check Out:

• Agentic Browsing in PageSpeed Insights: How to Prepare Your Site for the Age of AI Agents
• Topic Clusters: What They Are, How to Create Them and Why They Improve SEO
• Trust Signals for AI: What Convinces Agents to Recommend Your Brand
• Brand Mentions in AI Search: How to Turn Citations Into Real Traffic in 2026