AEO Best Practices

AEO content planning starts with a prompt set: the questions buyers actually ask AI assistants across awareness, consideration, evaluation, and decision. Those prompts get run against the engines to see who is cited today, which produces a gap list. Gaps are then ranked by commercial value, so comparison, alternatives, and pricing questions usually outrank generic definitions, with priority going to questions where the company can contribute original proof.

Start with prompts, then find the gaps

The prompt set replaces the editorial brainstorm. HubSpot built its program by mapping buyer questions across awareness, consideration, evaluation, and decision, then measuring visibility, share of voice, citations, and citation share against each one (HubSpot AEO case study). That map shows you which prompts already name you and which ones name a competitor in the slot you wanted. The most useful category is the third one: prompts the engine currently answers without any good source behind it, where the first credible answer published tends to get adopted and then repeated.

We run the same exercise before writing anything: 50 to 100 real prompts, scored across ChatGPT, Claude, Gemini, and Perplexity, with the current answer captured verbatim. The verbatim capture is what makes it worth doing, because watching an engine confidently describe your product in a competitor's framing produces a sharper content brief than any keyword tool is going to hand you.

How we rank the gaps

Not every gap is worth closing. We weight them roughly like this:

Definitions still get written early, because they're how an engine learns what you are and they're inexpensive to produce, though they rarely deserve more than the opening weeks of a plan. In our 90-day plans we typically front-load definitions and comparisons in the first three weeks, then spend the rest of the runway on the commercial prompts where a citation actually changes a deal.

The proof test

The last filter is the one that kills the most ideas. For each candidate question we ask what we can say that nobody else can, whether that's a number out of our own client data, a process we actually run on Monday mornings, or a case study with a client's name attached to it. When the honest answer is that we'd be paraphrasing what's already published, we don't write the page.

This is also where the format changes. Kevin Barber, our Head of AI Growth, has made the same argument about format: "AEO is substantially different than traditional SEO because instead of long-form pages, you're focusing on semantic chunks that answer questions in bite-sized components." A 4,000-word guide that buries its answer in section six is difficult for a model to extract from, so we take the same expertise and organize it as discrete question-and-answer units in an answer hub, which gives the engine clean passages it can lift without needing to reconstruct the argument.

Where search demand still fits

Prompts and keywords work as complementary inputs, and we run three streams together: the language sales hears on calls and in the CRM, the citation gaps from the baseline, and search demand data from Semrush to sequence what gets written first. Search volume is a decent proxy for how many people are asking a question in any interface, and it's the only one of the three with a long history behind it.

The blend is what makes the calendar defensible. CRM language keeps the content in the buyer's actual words instead of internal product language, while the citation gaps aim it at the questions where we're currently absent. Demand data is the sanity check that stops us from writing a genuinely great answer to a question three people a year ask. It's the same sequencing logic behind how we build content programs generally, applied to a set of prompts instead of a set of keywords.

Most existing HubSpot content can be upgraded in place. The work is moving the direct answer to the top of the page, rewriting headings as questions buyers actually ask, adding specific proof, and applying valid schema where the type genuinely fits. HubSpot supplies the measurement layer too, with AEO tooling in Marketing Hub Pro and Enterprise and an AI Referrals traffic source that connects AI-driven sessions to contacts and deals.

Upgrading pages you already have

A rebuild is rarely the starting point. Most of the blog posts and pillar pages in a mature HubSpot portal already contain the expertise; the structure just wasn't built for extraction. The retrofit is mechanical:

• Move the direct answer into the first 50 to 75 words, written so it stands alone if a model lifts it out of context.
• Rewrite decorative H2s as the questions buyers actually type.
• Make every paragraph self-contained, so a passage pulled from the middle of the page still makes sense without the paragraph above it.
• Add the specific number, process, or named case that you have and your competitors don't.
• Tighten internal linking so related answers form a cluster a crawler can traverse.

We typically get through 40 to 60 existing pages in a Launchpad engagement. HubSpot's modular templates are what make that pace possible, because a well-built answer module can be dropped across an entire content library instead of hand-edited page by page. The test we hold each retrofitted page to is one you can run yourself. Pull any paragraph out of the middle of the page and read it with nothing around it, then ask whether it still answers something a buyer would type into a chat window. About half the paragraphs in a mature blog fail that on the first pass in our experience, and most of them pass after a structural edit that never touches the underlying argument.

Schema is the layer that makes those facts explicit, and HubSpot's templating makes it maintainable at scale rather than something you hand-code into 200 pages and then abandon. We ship it through Schema Rocket on HubSpot builds, keeping the markup identical to the visible content and validated on publish.

What HubSpot gives you natively

HubSpot ships AEO capability inside Marketing Hub Pro and Enterprise, and also sells it as a $50 per month standalone product (HubSpot's AEO announcement). It tracks visibility across ChatGPT, Gemini, and Perplexity, suggests prompts based on your actual CRM data, and reports competitor visibility, sentiment, and citation analysis.

For $50 a month it's a straightforward yes on the measurement side, provided you're clear about the boundary between what it reports and what it changes. It will show you that a competitor owns the comparison prompt in your category, and closing that gap still means writing the comparison, building the page, and earning the corroboration that sits behind it. We use the tool as the input to a retrofit queue: the prompts where HubSpot says you're absent become the next batch of pages we rewrite.

Tracking AI-driven traffic and pipeline

HubSpot's analytics now classifies visits from ChatGPT, Claude, Perplexity, Copilot, Gemini, Meta AI, Mistral, Poe, and Grok as AI Referrals (HubSpot traffic sources documentation). That flows into contact and deal records: Original Traffic Source and Latest Traffic Source can both read "AI Referrals," with the platform domain in Drill-Down 1 and campaign in Drill-Down 2 (source property documentation). ChatGPT helps by appending utm_source=chatgpt.com to its search referral URLs (OpenAI publisher FAQ).

For a sales-led company this is the most useful thing HubSpot does for AEO, because it puts AI-sourced pipeline in the same reporting object as everything else. On other stacks we usually find the AI visibility data sitting in one tool while the revenue data sits in a CRM that has never heard of it, so answering "did AEO produce anything this quarter" turns into a manual join somebody has to redo every month, and in our experience it quietly stops getting done around month four.

There's one blind spot worth planning around. Ahrefs found AI assistants linked to its site in only 28% of brand mentions (citation study), so AI Referrals will always under-report your real exposure. We pair HubSpot's traffic data with prompt-level citation tracking to cover the mentions that never produce a click.

If you're not on HubSpot yet

The AEO work doesn't require HubSpot, though it's noticeably easier there, particularly for sales-led B2B SaaS teams that want AI-sourced contacts attributed automatically. As a Diamond Solutions Partner we've run this on both sides of that line, and when a platform migration is already on the roadmap for other reasons, folding the AEO structure into the rebuild costs far less than retrofitting the same pages twice.

AEO produces results in stages, each on its own clock. Crawler access and indexable answers can be corrected in days. First mentions and citations on specific, lower-competition prompts commonly appear within four to twelve weeks of publishing. A stable share of voice across several engines generally takes three to six months, because engines refresh their source pools on their own cycles while competitors keep publishing against the same questions.

What counts as a result changes the timeline

The word "result" covers at least four separate measurements, and each one moves on its own clock. We plan engagements against all four so nobody is surprised in week three when the dashboard is still mostly empty.

Technical eligibility is the only genuinely fast piece of the work, and it's easy to overlook because nothing in your analytics flags it. A blocked crawler will keep you out of an answer your content would otherwise have earned, and so will text that only renders in the browser, or a canonical tag pointing somewhere unhelpful, which is why Google's own AI features guidance treats normal Search eligibility as the entry requirement for AI Overviews and AI Mode.

Why we say 60 to 90 days

Most of our clients see measurable citation and recommendation activity within 60 to 90 days of completing the AEO Authority System setup, and that activity compounds instead of arriving all at once. The window tracks how the work sequences. The on-site foundation (answer hub, entity and schema markup, solution grounding pages) goes live in the first few weeks, while off-site corroboration has to be pitched, published, indexed, and then pulled into an engine's source pool, which adds weeks that are outside anyone's direct control.

Other providers publish similar windows. Single Grain cites initial citation improvements at two to three months with more substantial visibility at four to six, which lines up with what we see when a company starts from a low citation base.

What makes one company faster than another

Existing authority is the biggest accelerant, and it's mostly earned before the AEO engagement starts. Semrush's 1,000-domain analysis found a 0.65 correlation between backlink authority and AI mentions, with quality mattering more than raw volume, so a company with real domain authority and an existing content library gets picked up faster than a company with neither. This is also why we say AEO layers on top of SEO rather than replacing it.

The fast public cases follow that pattern. Sandler's AEO launch produced 8,000 new visitors and 4,000 ICP-fit prospects inside a few weeks, and the reason it could move that quickly is sitting in the correlation data above: they arrived with the authority and the content library already built, so the engines had plenty to retrieve from day one.

A few other variables move the date around, and they're worth being honest with yourself about before you set a board expectation. Category demand comes first, because if buyers already ask AI about your space you're competing to change an answer that exists, which is faster than teaching the engines that the question is worth asking at all. Crawlability comes next, and a site the engines can already read saves you the first few weeks outright. Then there's your bench, since the answers that actually get cited tend to carry a specific number or a defensible position that only your team can supply. Publishable proof is the slowest input of all, because authentic reviews and third-party mentions can't be manufactured inside month one.

How to read your first 90 days without fooling yourself

Judge movement against a fixed prompt panel that you rerun on a schedule, since a single spot check on a Tuesday tells you almost nothing. Engine source selection is genuinely volatile: Semrush measured roughly 120% average prompt-coverage change among ChatGPT's top 100 source domains over three months, so a citation appearing or vanishing in a given week often has nothing to do with your last content edit.

Annotate model updates alongside your own launches so you can tell which of the two moved the number, and score each engine separately, since they disagree with each other far more often than they agree. All of that needs a documented starting point to be worth anything, and if you don't already have one, our baseline score is free and will hand you a dated citation gap report to measure the next 90 days against.

AEO requires real internal time even when an agency handles research, writing, schema, outreach, and monitoring. The company still supplies positioning, customer language, proof, product accuracy, expert review, and approvals, because those inputs cannot be sourced externally. A typical managed engagement asks for a heavier first month, roughly 10 to 18 hours across all internal contributors, then 5 to 10 hours per month once the review cadence settles.

Where the hours actually go

Internal time concentrates in the inputs that only you have. Here's the planning estimate we give clients before kickoff, split by role rather than by person, since one founder often wears several of these hats.

The SME interview is the line item people underestimate and the one that decides whether the content is worth publishing. Engines reward answers that contribute something checkable, and checkable material lives in your experts' heads: the number they'd quote from memory, the process they'd sketch on a whiteboard if you asked. An hour of your VP of Product on a call typically produces material for several answers, which is a good trade for both sides.

Customer proof carries its own cost. Reviews at scale, expert quotes, and case evidence need your customer relationships and your permission to use them, so budget time for the introductions rather than assuming the agency can conjure third-party credibility on your behalf.

What automation genuinely removes

Measurement is the part of AEO that gets dramatically cheaper with the right setup. Sandler cut 16 hours per week of manual signal stitching down to a 15-minute daily dashboard check after consolidating its AI visibility reporting, which is a fair picture of what good tooling does to the reporting workload.

That's the trade we're aiming for in the monitoring layer of our system: your team looks at a citation dashboard and a competitive positioning summary instead of manually rerunning prompts across four engines and pasting screenshots into a deck. Prompt reruns, citation tracking, competitor movement, and gap identification are mechanical enough that they don't belong on your marketing lead's calendar at all.

Content production, technical implementation, publisher outreach, and schema maintenance can also be handed off. What stays with you is judgment about whether a claim is true and whether you'd defend the positioning in front of a buyer, which is not something you can outsource to anyone who doesn't work at your company.

Why a very low client-time promise is a warning sign

A provider who promises you'll barely be involved is usually describing a process that doesn't need you, and the output tends to show it. Content assembled without your input ends up accurate in the way a Wikipedia summary is accurate, which means it reads like every competitor's version of the same page, and once four vendors have published essentially that paragraph, an engine's choice of which one to cite comes down to authority instead of substance.

The quieter failure runs the other way. Some programs push agency work back onto the client, so you end up drafting, sourcing, and building schema yourself while paying a retainer for a dashboard. Before you sign, ask for the hours broken out by role and by month, and ask what the plan is for month four when your marketing lead is on vacation and the review queue backs up. Our published AEO pricing separates the one-time on-site foundation from the ongoing authority work partly because the two make very different demands on your calendar.

The practical version of this is a division of labor we set at kickoff. We own the structure, the markup, the outreach, and the monitoring, and the material that goes inside all of it has to come from people who work at your company. When a client's SME hours slip for two months running, the citation curve slips with them about a quarter later, and that lag is what we point at when somebody asks why we're so persistent about a recurring 45 minutes on a product manager's calendar.

AI referral tracking works on two layers. Analytics platforms identify visits arriving from assistants: HubSpot classifies traffic from ChatGPT, Claude, Perplexity, Copilot, Gemini, Meta AI, Mistral, Poe, and Grok as AI Referrals, and OpenAI appends utm_source=chatgpt.com to ChatGPT search referral URLs. Because many assistants and privacy features strip referrers, and because most AI mentions never produce a click at all, referral analytics must be paired with prompt-level citation tracking.

Setting up AI referral reporting in HubSpot

HubSpot has a native AI Referrals traffic source in the traffic analytics tool, so there's no custom channel grouping to build. It recognizes visits from ChatGPT, Claude, Perplexity, Copilot, Gemini, Meta AI, Mistral, Poe, and Grok, and it exposes the specific platform in Drill-Down 1 with campaign parameters in Drill-Down 2, per HubSpot's traffic source documentation.

The part worth wiring up carefully is contact and deal attribution. Original Traffic Source and Latest Traffic Source can both be set to AI Referrals, with the assistant's domain in Drill-Down 1, according to HubSpot's source property documentation. That gives you a list view or report of every contact whose first touch came from an assistant, which is the closest thing to real AI-sourced pipeline reporting available in a CRM right now. We usually build three saved reports at the start of an engagement: AI-referred sessions by platform, contacts created with Original Source of AI Referrals, and deals with any AI Referrals touch in the path.

If you're on GA4 instead, you'll be building a custom channel group with a regex against the known assistant referrer domains, and maintaining that list yourself as new engines appear.

What UTMs and referrers actually capture

ChatGPT gives you the cleanest signal of the major engines. OpenAI states that referral URLs from ChatGPT search automatically include utm_source=chatgpt.com, which survives most analytics setups and gives you an unambiguous filter.

The rest of the ecosystem is uneven. Some assistants pass a referrer and some pass nothing at all, while plenty of answers get rendered in a way that means the user never leaves the chat window in the first place. Copied links, in-app browsers, and privacy features strip parameters on the way through, so the AI Referrals number you see is a floor, and often a badly understated one.

Why referral analytics undercount, and what to do about it

Most AI brand exposure never becomes a link at all. Ahrefs found that AI assistants linked to its site in only 28% of brand mentions on average, so for every mention that hands you a measurable click there are two or three that hand you nothing your analytics can hold onto. A buyer can read a recommendation of your product, form an opinion, and show up three weeks later as a direct visit that your dashboard has no memory of.

Tracking the answer itself is what closes that gap. We run a fixed prompt panel across ChatGPT, Claude, Gemini, and Perplexity, rerun it on a schedule, and record whether you appeared, whether you were linked, how you were framed, and who else got named alongside you. That gives you mention rate, citation rate, share of voice, and brand framing accuracy, none of which referral analytics can reach. You can see what the monitoring looks like in practice on our demo of how AEO works, and the same tracking sits inside our AEO Genie tooling.

Neither layer stands on its own. Prompt tracking will tell you the engines know you and recommend you without telling you whether that produced a single dollar, and the CRM data closes that gap once the referred visitors start converting. A program judged on referral analytics alone almost always looks worse than it is, because that's the layer missing most of the exposure.

An early-stage SaaS company should expect visibility to move well before pipeline does. AI referral traffic stays small in absolute terms for most sites, often a fraction of a percent of total visits, though it tends to convert unusually well because the assistant has already filtered the visitor. Realistic first-90-day outcomes are crawl eligibility, baseline prompt coverage, accurate brand framing, and first citations on high-intent questions.

The traffic numbers are small, and that's the normal case

AI referral volume is a rounding error on most sites today. Semrush's channel mix analysis across 50,000 sites found that AI traffic grew 66% in 2025 but still accounted for less than 0.15% of total visits. If you launch AEO expecting a traffic chart that looks like paid search, you'll be disappointed in month two and you'll shut the program down right before it starts compounding.

Volume is also the wrong first metric for a startup. You're buying position in an answer that gets composed whether or not you participate, and the cost of being absent from that answer is invisible until a competitor becomes the default recommendation. What we see across client dashboards is that the absence shows up first as an unremarkable quarter, with the same lead volume and the same pipeline as always, while buyers who would once have found you through a comparison search are now handed a shortlist you aren't on and never bounce off your site at all.

Why the small traffic is worth having

Visitors arriving from an assistant have usually been pre-qualified by the assistant. Ahrefs published the sharpest public example: 0.5% of its visitors came from AI and produced 12.1% of signups, a roughly 23x conversion ratio against traditional organic search. That's one company, one product, an unusually strong brand, and a free-tier signup path, so it belongs in your thinking as evidence that AI traffic can convert exceptionally well, and nowhere near your forecast spreadsheet.

The mechanism is intuitive enough. Someone who asks an assistant to compare tools, reads a synthesized answer, and then clicks through to your site has already done the filtering that a search visitor does after landing. Buyer behavior backs this up: G2's survey of 1,076 B2B software buyers found 51% now start research with an AI chatbot more often than Google, and 71% use one somewhere in the process.

What a realistic first 90 days looks like

Hold us to foundation and coverage in the first quarter, since neither citations nor pipeline can exist before those two are in place.

Brand maturity moves this whole table. The fast public AEO cases come from companies like HubSpot and Sandler, both of which arrived with recognized brands and large existing content libraries. A Seed-stage company starting with no citations and little third-party evidence is running a different race, and borrowing those timelines as a forecast sets your board up for an uncomfortable conversation in Q2.

What you need in place for any of this to work

Engines can only repeat what's already verifiable about you. If your positioning is still moving week to week, or you don't yet have customers who'll go on record, we'd tell you to spend a quarter on that before you spend anything on AEO, because the program has very little to work with until you do. We fit Seed through Series C companies on HubSpot that have a real offer and a measurable goal, which is a deliberately narrower statement than "we work with startups." You can see how we've handled that stage in our SaaS work and the outcomes we've published, where the headline numbers (740% organic growth for Qualio, 300% more organic leads for The Predictive Index) came from multi-year programs across web and organic, so they tell you we can move a number without telling you what AEO on its own does in a quarter.

Stopping AEO does not erase existing visibility. Published answers, schema, reviews, and third-party mentions remain retrievable, so engines can keep citing them well after the work stops. The change shows up in relative position, because competitors keep publishing and models periodically refresh which sources they draw on. Decline is gradual, unfolding over months, and a scaled-down maintenance program preserves most of a hard-won position at much lower cost.

What keeps working after you stop

The assets you paid for are published artifacts, so they keep doing their job without you. Answer pages remain on your domain where crawlers keep finding them, and the schema we deployed keeps describing your brand facts to anything that reads it. Reviews live on the platforms buyers already check, and any expert article we placed sits on the publisher's site, inside the evidence pool engines draw from. We've seen brands hold citations on their core prompts for months after production paused, which is why we don't tell anyone their visibility evaporates the day they stop paying.

The same is true of your entity footprint. Directory listings, consistent brand facts, and corroborating mentions across the web do not expire on a billing cycle. If we built you an answer hub and entity structure through our AEO Authority System, that structure survives a pause, though it stops growing while you're paused.

What actually erodes, and how fast

Your position is measured against a moving field. Semrush tracked ChatGPT's top 100 source domains over three months and found the average change in prompt coverage was roughly 120%, meaning the set of sources an engine leans on for a given question churns significantly in a single quarter (Semrush AI visibility trend update). Standing still in that environment costs you ground, because a competitor publishing a better answer to a high-intent prompt can displace you without anything on your side breaking.

Other drift is quieter. Facts go stale, so a model can keep recommending you on the basis of pricing or positioning you've since changed, and coverage gaps open as new questions enter the category with no page of yours addressing them. The recommendation itself can also go wrong in ways your analytics will never surface, describing you as something you aren't, which we only ever catch by rerunning the prompt panel.

Honestly, nobody has published a clean decay curve for a paused AEO program, and any vendor quoting you one is guessing. What we can say from watching client dashboards is that month one usually looks flat, and the movement you notice tends to appear in months two through six.

A maintenance mode that actually makes sense

If budget is the reason you're asking, there's usually a version of this that costs a fraction of the full program, because most of the spend went into building the foundation, and keeping that foundation accurate is comparatively cheap. Here's how we'd usually split it:

A maintenance posture like this holds your accurate brand framing in place and gives you a heads-up if a competitor starts picking up the recommendations you used to get. It also keeps the restart cheap, because if you come back in six months the foundation is still standing and the work picks up roughly where it left off.

Drift is also something you can keep an eye on without a retainer. Scoring your visibility across ChatGPT, Claude, and Gemini costs nothing, and running it every month or two through a pause will show you which prompts you've quietly stopped appearing in, which is usually the first thing anyone notices.