The cheapest failure in AI search takes an afternoon to find and far longer to undo: a page no engine can fetch. The most expensive one is a quarter of content spend aimed at a problem nobody measured first. For a team working with a small budget and no dedicated writer, the order of operations decides more than the tactics do, and the order runs cost-to-check before cost-to-fix.
Here is the sequence in three steps. Verify that an engine can retrieve your pages at all, then record a baseline on five to ten buyer questions so you can tell later whether anything moved. Spend what is left on a short, ordered set of evidence pages, starting with the ones a buyer checks while building a shortlist. The rewrite you were about to commission waits, and the reason is in the published evidence rather than in taste.
The prize for getting this right is unusually large for a company with no brand recall. 51% of B2B software buyers in G2's survey now start research with an AI chatbot more often than with Google, G2 found AI chatbots and review sites are now the two biggest influences on B2B shortlists, and in the same survey a third said an AI chatbot led them to choose a vendor they hadn't considered. That is a door that opens from outside the buyer's plans, the only kind a startup can walk through.
Is authority the right word for what you are buying?
Use it carefully, because no public source establishes that generative engines compute an authority score for your domain and then spend it on your behalf. Google's own documentation on its AI features states there are no extra requirements or special optimizations to appear, no new machine readable files or markup, and that pages must be indexed and snippet-eligible. That is a fetch-and-eligibility condition rather than a reputation dial.
What replaces the single dial is a set of separate questions that can move in different directions at once. The IAB's 2026 framework splits AI-era visibility into presence, prominence, portrayal and persuasion. A team with one contractor cannot move four measurements. It can pick one and defend it.
What should the first afternoon buy?
A fetch test on your ten most commercially important pages, done by hand, with no vendor involved. Most AI crawlers do not execute JavaScript: crawler research Vercel ran with MERJ documented that these agents read raw HTML, so a client-rendered page arrives as an empty shell. Cloudflare scanned the 200,000 most visited domains for agent readiness and found the standards built for agents barely adopted: only 4% of sites declared AI usage preferences in robots.txt, and 3.9% passed Markdown content negotiation.
Request each page the way a crawler would, without a browser, and read what comes back. Confirm your robots rules allow the agents you want by name, because they are declared separately from Googlebot; Cloudflare measured GPTBot's share of crawler traffic rising from 2.2% to 7.7% in a year. Confirm the page is indexed, because Google ties AI feature eligibility to indexing.
This step comes first for an economic reason. It costs a few hours, it produces a yes or no, and a no turns every downstream dollar into zero.
Why record a baseline before you produce anything?
Because AI answers vary enough between runs that a single reading cannot support a claim in either direction. SparkToro and Gumshoe's research found AI engines are highly inconsistent when recommending brands, and Search Engine Land's coverage of the same study notes the recommendation lists rarely repeat exactly. Publish first and you have no honest way to attribute anything that happens afterwards.
A 2026 variance-components study found run-to-run variation large enough to swamp real differences in small samples, and Ronald Sielinski's work on quantifying uncertainty in AI visibility makes the same argument with confidence intervals. Our own floor is several weeks of answers per question, across phrasings and engines; neither study sets a number, and one run is not a measurement.
Keep the question set small and commercial: five to ten questions a real buyer asks before choosing, run on at least two engines, recording who was mentioned, whose page was cited, and who was recommended. Where those questions come from is its own problem, and the ideation piece covers it.
Should the budget go into rewriting your existing pages?
No, and this is the correction that saves a constrained team the most money. C-SEO Bench, published at NeurIPS 2025, evaluated conversational-SEO rewrite methods and found most of them largely ineffective and often harmful to citation rank, with gains shrinking as more sites adopted them. The benchmark tested those methods across two tasks and six domains, so the null result is broad rather than a quirk of one dataset.
That sits against the GEO study from KDD 2024, which reported that adding statistics, quotations and citations lifted visibility in its benchmark, while keyword stuffing was among the least effective methods. C-SEO Bench retested those same methods and found no significant gains from them, so the two results do conflict. Rewriting adjectives is not adding evidence, and neither is a proven citation lever.
Some persuasion language costs you outright. Research on scarcity and exclusivity framing found it measurably reduces how often an LLM recommends a product. Rewrites look cheap because the pages already exist, and on the published record they are the weakest use of a small budget.
Which ten pages should a new B2B startup publish first?
The pages a buyer opens while deciding, published in the order the decision uses them and scoped to one narrow territory where your own experience is the best answer available. A small team cannot fund coverage, so it funds specificity: the question your customers actually asked last quarter, answered with what you saw. The territory decides what the definition, how-to and common-mistakes pages are about, and the ten-page set is how you fund it.
Two constraints shape the pick. Profound's citation research found more than 4 in 10 AI citations point to sources brands do not control, so a question already settled by editorial roundups and community threads is one you cannot enter cheaply from your own domain. Seer found 87% of SearchGPT citations matched Bing's top results in a 500-citation sample. That overlap does not prove incumbency causes citation, but a term incumbents already rank for means competing against their retrieval as well as their content.
Both conditions are testable in an hour: ask the question in an engine and see whether review platforms or forum threads carry the citations, then run it through a plain web search and see whether long-established vendors hold the first page. A question that fails neither test is worth funding. Semrush's AI Visibility Index is built on 126 million AI search prompts, the scale of volume the broad category questions already attract. If your territory is original evidence you already hold, the data-source piece is the deeper treatment.
Shortlists form early: 6sense found the typical buying group evaluates about five vendors and nearly four out of five are named to the shortlist on day one. Most of these pages are about your own product, where you hold first-hand evidence a review site has to borrow.
A comparison page against the vendor buyers name most often, with the trade-offs stated plainly. The comparison-page guide covers structure.
An alternatives page for that same vendor, where you are one honest option among several rather than the winner of every row.
A pricing page with the numbers in server-rendered HTML, or real ranges and the factors that move them.
One case study with a named customer, a dated starting point and a measured result you can source.
A definition page for the narrow problem you picked as your territory, opening with a short answer that can be quoted whole.
A how-to for the core job that a reader can follow without buying anything.
An integration page for the tool most of your customers already run, with setup steps and limits.
A who-it-is-not-for page that names the buyers who should pick someone else and why.
A common-mistakes page drawn from what you watched customers get wrong, not from other vendors' lists.
A proof page that lists each claim you make anywhere on the site next to the source behind it.
The first five carry the shortlist moment, so ship them before six through ten. Google's guidance for review content lists 14 recommendations, including first-hand evidence and stated drawbacks; it was written for reviews, but borrow that standard for comparisons. A relevant page can still miss the proof, and this list is ten chances to avoid that.
What should you deliberately not do yet?
Do not build the cluster. Interlinked topic architecture is a real body of practice and it has its own argument here, but it presupposes volume you cannot fund this quarter, and no published study shows that a content structure by itself causes citations. The ten pages above are an evidence set, not a cluster.
Do not buy a scoring tool at all. A composite score compresses mentions, citations and recommendations into one number that cannot be acted on. And do not expect a single asset to carry a deal, since the same 6sense research found buyers already know 75% of the vendors on their shortlist by the time it is filled.
Do not chase the broad category term either. Ahrefs found 38% of AI Overview citations rank in the organic top 10, so rank alone is a weak gate. That does not make the broad term affordable, only differently expensive. Say the trade out loud to whoever approves the spend: findable on a small set of questions, absent from the general one, by decision rather than by surprise.
How does Trovance fit a constrained sequence?
Trovance runs the second step as a system rather than an afternoon of manual prompting. You define the buyer questions worth tracking, and it sweeps them across AI engines every week, preserving each answer as a snapshot with its full context: who was mentioned, whose page was cited, who was recommended, and which sources carried the answer. That preserved record is the baseline the rest of the sequence depends on.
Because the snapshots accumulate, share of answer becomes something you can compare over time. A small team can watch a handful of tracked questions and see whether a pattern is a real absence or the run-to-run variance the research describes, which is the difference between spending on a problem and spending on noise.
The production side is deliberately narrow. The Brain keeps the claims you are entitled to make with the proof behind each one, and drafts such as a comparison page or a case study are written in your voice from your voice guidance and research sources. Recommended actions name where the record shows you absent, and a person decides whether that absence is worth an asset. Nothing publishes without that review.
What Trovance will not promise is a citation, a ranking or a recommendation on a schedule. The engines are probabilistic, the sources shift, and your competitors publish too. What it does instead is close the loop: the next analysis cycle reruns the same tracked questions after your page ships, so you can verify whether the answer moved rather than assume it did.
What should you do this week?
Spend two hours on the fetch test and fix whatever it exposes, because nothing downstream survives a negative result there. Spend the next block writing down five to ten buyer questions and asking each in two or three phrasings on two engines, weekly for a few weeks, recording mentions, citations and recommendations separately. Then draft page one of the ten, the comparison against the vendor your baseline names most.
Write the steps down as a budget: hours against the fetch test, a fixed number of runs against the baseline, and the remainder against the first five pages. A plan that names amounts survives a bad week, because the first thing a squeezed quarter cuts is the measurement nobody costed. Be blunt about the timeline too: retrieval fixes can surface once engines re-fetch the page, earned third-party presence takes longer, and nobody can honestly guarantee an outcome inside the measured variance.
If you would rather have the baseline running than assembled by hand, start a free Trovance trial and let the first cycle tell you which step you are actually stuck on.
Check before you spend
AI crawlers don't run your JavaScript - the fetch failure that invalidates every later dollar.
Technical SEO for AI search - the mechanical checklist behind step one.
The 10-minute audit of what ChatGPT tells buyers about you - the fastest possible first reading.
How to measure AI search visibility without one score - what to record instead of a composite number.
Spend the rest deliberately
How to build a competitor comparison page - the first of the ten pages, done honestly.
Become a data source AI cites - the deeper case for original evidence over commentary.
Content ideation for AI visibility - where the five to ten buyer questions come from.
Human in the loop content review - how much review a small team actually has to fund.
