At a few deals a year, the usual case for AEO fails and a different one holds. The traffic case fails because AI assistants send small volumes that mostly cannot be attributed, and because a rise in direct traffic is not the proof it is sold as: Kevin Indig measured total direct traffic up 140.9% across a filtered sample while engaged direct traffic stayed flat (Growth Memo, October 2026). The case that holds is shortlist inclusion. In a 200-person Wynter survey of software buyers analysed by Kevin Indig (Growth Memo, November 2026), 79% arrived at the AI chat with a vendor already in mind, and 62.7% of those added another vendor through it. If you close eight deals a year, one added shortlist place is not a marginal gain, it is the return. So judge the programme on whether you appear on your buying questions, and build the business case on that rather than on sessions.
AEO is often more worth it at a few deals a year than at a few hundred, and only if you judge it on shortlist inclusion rather than on traffic. The arithmetic that makes the usual case for this work, visits multiplied by a conversion rate, breaks at low volume. The arithmetic that replaces it is cruder and more honest: what is one deal worth, and what would it be worth to be named when a buyer asks an assistant who to consider.
Two things make low volume the favourable case rather than the awkward one. A single contract can cover a year of work, so the programme does not need a measurable traffic channel to pay for itself. And the questions your buyers ask tend to be narrow, which means the number of prompts worth winning is small enough to actually win. A firm selling to a hundred thousand buyers is competing for broad category answers against incumbents with a decade of domain authority. A firm selling to four hundred is competing for a dozen specific questions.
What makes low volume genuinely hard is proof. At eight deals a year you will never have a statistically meaningful before and after, so the programme has to be defensible on leading indicators and on the judgement that being on the shortlist matters. Anybody who offers you a revenue attribution model at that volume is selling you a spreadsheet, not a measurement.
The rest of this page sets out why the traffic case fails, what the survey evidence says about shortlists, how to write the business case without inventing a conversion rate, and the situations where the honest answer is that this is not your problem to solve yet.
The traffic case for AEO fails at low deal volume because the channel is small, badly labelled, and easy to mistake for something else, so the one number a business case wants is the one number nobody can produce. It fails for a measured reason rather than a philosophical one.
Kevin Indig tested the most common version of the claim, that a rise in direct traffic is hidden AI traffic. His method was 12 Google Analytics 4 properties compared across August and September 2026, excluding homepages, pricing, login and other high-direct pages, plus 632 comparisons of weekly AI visibility against direct and branded search at lags of zero to six weeks. Over January to September, total direct traffic rose 140.9% while engaged direct moved by -1.0% on monthly counts and +2.3% on average daily volume, and one site logged 409,000 direct sessions of which 198 were engaged, according to "Does AI drive more direct traffic?" (Growth Memo, 19 October 2026).
His correlation tests are the more interesting half. Two brands showed relationships of 0.72 and 0.47 between AI visibility and direct traffic, and a shifted-timing control found that several wrong alignments produced equally strong or stronger correlations, with each site's relationship appearing in only one half of its timeline. His conclusion was that he could not detect hidden AI traffic without self-reported attribution. That is a practitioner study with a small usable sample, and it is the most careful public attempt at the question so far precisely because it reports the control that undermined its own positive results.
For a firm closing a few large deals, this is not a reason to skip the work. It is a reason to stop promising the wrong evidence. Our guide on whether AI search actually sends traffic to your website covers how small the labelled part of the channel is, and whether a direct traffic spike is hidden AI traffic works through the test on your own data.
AI search changes a buyer's shortlist mainly by adding to it, and that is the finding that matters most when you sell a small number of large deals. The best stated-method evidence available is a survey of 200 participants with software in their industry profiles, run by Wynter and analysed by Kevin Indig in "How software buyers use AI" (9 November 2026).
The survey found that of the 200 respondents, 158, or 79%, had at least one vendor in mind before opening the chat. Of those 158, 99, or 62.7%, said they added another vendor through it. Of the 126 who reported choosing a vendor, 78, or 62%, said it appeared in the first answer, and 43 said it appeared only after follow-up questions. The follow-ups themselves were pricing at 57.5%, comparing vendors at 54.5%, integrations at 46% and setup at 40%. After the chat, 72.5% visited vendor websites and 29% searched Reddit.
The limitations travel with the numbers. The survey is self-reported and unweighted, Indig states it is not representative of all B2B buyers, recalled prompts may be wrong, a reported demo does not confirm a purchase, and he discloses that he is an investor in Wynter. Treat it as the first careful description of the behaviour rather than a population estimate.
Read against a low deal count, the useful sentence is that four buyers in five already know who they might buy from, and roughly three in five of those come out of the conversation with a name they did not walk in with. That is the slot you are buying. It also tells you which pages matter: pricing, comparison, integrations and setup were what buyers asked next, in that order, so a site that answers the category question and goes quiet on the four follow-ups loses the thread. Whether B2B buyers use these tools at all is covered in our guide on how B2B buyers research vendors with AI.
Build the business case on one added shortlist place rather than on a funnel, and state the assumption in the open where your CFO can argue with it. At low deal volume the honest model has three inputs and no conversion rate: what a deal is worth, how many of your buying questions you currently appear on, and what a year of the work costs.
Write it as a break-even rather than a forecast. If a year of the programme costs less than a fraction of one deal, the question is no longer whether the return is provable but whether the mechanism is plausible, and the survey evidence above is what makes it plausible. That framing also survives contact with procurement, because it does not claim a number nobody can audit. Our guide on getting budget approved for AI search visibility covers the rest of the internal path, and what moves an AEO quote up or down covers the cost side.
Then set the reporting expectation in the same document, because the commonest failure at this volume is a programme judged in month four against evidence that cannot exist until month nine. What you can show early is appearance: which of your buying questions name you, on which engines, and whether that set is growing. What you cannot show early is revenue.
| Deals a year | What the business case should rest on | What to report monthly |
|---|---|---|
| Under 10 | Break-even against a fraction of one deal | Appearance on named buying questions, plus self-reported source on every lead |
| 10 to 50 | Break-even, with pipeline influence as a secondary read | Appearance, share of voice against named competitors, lead source |
| 50 to 500 | Pipeline contribution, with attribution caveats stated | The above, plus labelled assistant sessions as a floor |
| Over 500 | Channel economics, treated like any other demand channel | The above, plus cohort conversion |
One rule holds across every row. Put a free-text question about where someone first heard of you on your own form, and read the answers. It is the only instrument in this list that catches the buyer who read your name in an answer and never clicked.
Measure appearance, measure it on the same questions every time, and keep mentions and citations as two separate numbers, because at a few deals a year every outcome metric you own is a sample of one. Appearance is the only thing in this chain you can read repeatedly and compare.
Keeping mentions apart from citations matters more than it sounds. BrightEdge found ChatGPT mentioning brands 3.2 times more often than it cites them, with underlying counts reported as 2.4 brand mentions against 0.74 citations per prompt, in its analysis of ChatGPT brand mentions against citations (14 August 2025, read 17 November 2026). It is one vendor's dataset with no published method, so treat the ratio as indicative. The direction is the point: a programme that counts links undercounts the thing a low-volume seller actually wants, which is being named. Our guide on mention rate against citation rate works through which to track.
Then size the noise before you read any movement. BrightEdge's volatility analysis put the stability inflection near 50 citations, where weekly volatility drops from 50% to 8%, and found frequently cited domains moving about 0.7% week to week against more than 50% for sporadically cited ones, in "AI Search Engine Citation Volatility: The 70x Stability Gap" (23 October 2025, read 17 November 2026). A small firm in a narrow category is on the noisy side of that line by definition, so repeated runs are the answer rather than a single check.
Be clear about what an improvement score is and is not. We publish two anonymised early case studies in which one client's audit score went from 17 to 68 and another more than doubled, and the page says plainly that the audit score is an input and share of voice is the outcome. Holding that line is what keeps a low-volume programme honest, because an input metric is the only one that will move fast enough to look good.
AEO is not worth it at low deal volume in four situations, and all four are worth checking before you spend anything. None of them is about company size.
The first is an absent question. If your buyers do not ask an assistant anything your category answers, because they buy through a framework agreement, a distributor relationship or a procurement list that predates the conversation, then the mechanism has nothing to act on. Our pages on whether AEO works if nobody searches your category and selling through distributors cover both cases.
The second is an immovable site. If nobody internally can publish or edit a page within a quarter, the page half of the work cannot happen, and what remains is off-site work that a different kind of firm does better. The third is an unresolved accuracy problem on your own basic facts, which is cheaper to fix directly than to out-rank. The fourth is timing: if a rebrand, a migration or a redesign is already scheduled, the sensible order is to let it land first, as our guide on waiting until after a website redesign explains.
Where none of those apply, the cheapest next step is a baseline rather than a decision. A free visibility assessment samples your buying questions across ChatGPT, Perplexity, Google AI Overviews and Claude and reads your public pages the way a crawler does, which answers the one question the business case turns on: whether you appear at all today. If you already appear on the questions that matter, a programme has less to add than a vendor will tell you, and that is a useful thing to find out for free.
Founders and marketing leads at firms closing a small number of large deals ask these five questions most often.
There is no deal count that settles it, because the deciding number is what one deal is worth against a year of the programme, not how many you close. A firm closing six deals at a large contract value has an easier case than one closing six hundred small ones, since a single added shortlist place can cover the year. What rules it out is not low volume but an absent buying question: if nobody is asking an assistant anything your category answers, the mechanism has nothing to work on.
Only by asking the buyer. Kevin Indig tested whether AI visibility could be detected in analytics across 12 properties and 632 timing comparisons and concluded he could not detect hidden AI traffic without self-reported attribution. A free-text question on your own form about where someone first heard of you is the one instrument that works, and at a few deals a year you can also simply ask on the call.
At low deal volume it is worth more than the click, because the buyer who reads your name in an answer and types it into a browser later is the whole outcome. BrightEdge measured ChatGPT mentioning brands 3.2 times more often than it cites them, so a programme that counts only citations is undercounting appearance by construction. Counting mentions and citations separately is the fix.
Usually yes, and the pilot should be scoped to the handful of questions your sales team actually hears rather than to a page count. A baseline on those questions tells you whether you appear at all, which is the single fact the business case turns on, and it costs a fraction of a year of work to find out.
That you spend a year improving pages, appear in more answers, and cannot prove any of it turned into revenue, because at a few deals a year the sample is too small for any statistical claim and the attribution path is broken by design. That is a real risk and it should be stated in the business case rather than discovered in month nine.
A free visibility assessment samples your buying questions across ChatGPT, Perplexity, Google AI Overviews and Claude, and reads your public pages the way a crawler sees them.
Request a free visibility assessment →