Your Homepage Isn't an Agent Brief

Nobody asks an AI assistant "tell me about Acme." They ask which vendor handles multi-entity consolidation under $2,000 a month with a Snowflake connector.

Your homepage cannot answer that. It was never built to.

It was built to make a human who already knows your name feel like they found the right company, and it does that job with tools that the research says actively work against you when a machine is reading. The instinct to audit the homepage for AI-readiness is understandable and almost entirely misplaced.

What is an agent actually doing when it shortlists you?

It is answering a constraint-shaped question, not a brand-shaped one. The agent receives something like "compare tools that do X for a team of Y at Z budget," retrieves passages that appear to contain those specifics, and assembles a comparison. Your company enters that comparison only if some page somewhere contains a retrievable answer to the constraint.

This is a different retrieval event than the one your marketing site was designed around. A human arrives at your homepage having already selected you. An agent never selects you first. It selects the question first, then finds whoever answered it.

C-SEO Bench, presented at the NeurIPS 2025 Datasets and Benchmarks track, tested this directly. One of its two tasks was product recommendation. Across six domains, the researchers found that traditional retrieval-side improvements, the unglamorous work of being the source that best matches the query, had a significantly larger effect on citation ranking than any of the content-modification tactics the field promotes. The page that contains the answer wins. The page that describes the company does not.

Why doesn't your homepage answer the question an agent is asking?

Because homepages are written in abstractions and agents retrieve on specifics. Go read your own. Count the sentences that contain a number, a named integration, a price, a supported standard, or a customer outcome with a denominator. On most B2B homepages the count is close to zero, and that is not a failure of the copywriter. It is the format working correctly for its intended reader.

A homepage compresses. It says "built for modern finance teams" precisely so it does not have to say "supports 14-entity consolidation with a Snowflake connector on the $1,900 tier." The compression is what makes it readable to a human in six seconds. It is also what makes it useless as evidence.

I ran this count on Trovance's own draft homepage while writing this piece and found four qualifying sentences on a page of roughly forty.

We are not exempt from the pattern, and the fix was not to rewrite the homepage. It was to accept that the homepage has a job it does well and build the specific pages elsewhere.

Does persuasion language help or hurt when a machine is reading?

It measurably hurts, and this is the finding that should change how you think about the problem. Researchers testing cognitive-bias framings in product descriptions found that social proof consistently boosted recommendation rate and ranking, while scarcity and exclusivity framing reduced visibility, published at EMNLP 2025.

Sit with which of those a homepage is made of. Limited availability. Join the select few. Only for teams serious about growth. Book before slots fill. The entire scarcity-and-status toolkit that reliably moves human buyers is, in controlled testing, a drag on machine recommendation. Meanwhile social proof, the one bias that behaves like evidence because it carries counts and named parties, is the one that helps.

That is not a coincidence. Social proof is the only item in that list that contains a verifiable fact. The rest are emotional framings with nothing to retrieve.

The pattern holds elsewhere. The GEO study presented at KDD 2024 found that rewriting content in a more persuasive and authoritative register produced no significant improvement, scoring 21.3 against a 19.3 baseline, while adding statistics, quotations, and citations produced 30 to 40% relative gains. Sounding convincing does not work. Being checkable does.

Can an agent even read your homepage?

Often not, and this is the least glamorous item on the list with the largest consequences. In the server-log study Vercel ran with MERJ across their network and two independent job-board sites, none of the major AI crawlers rendered JavaScript. That list covers OpenAI's GPTBot, OAI-SearchBot and ChatGPT-User, Anthropic's ClaudeBot, Meta's external agent, ByteDance's Bytespider, and PerplexityBot.

They do download the JavaScript. ChatGPT's crawler fetched JS files in 11.50% of requests and Claude's in 23.84%. Neither executed them. If your hero copy, pricing table, or integration list is injected client-side, the crawler received a shell.

The exceptions matter for scoping: Gemini inherits Googlebot's rendering infrastructure and processes JavaScript fully, and AppleBot renders through a browser-based crawler. So a page can rank on Google, get read by Gemini, and be blank to ChatGPT at the same time. That is a real state your site can be in right now without any dashboard telling you.

Two more findings from that dataset are worth carrying:

A third of AI crawler traffic hits nothing. ChatGPT's crawler spent 34.82% of its fetches on 404 pages and Claude's 34.16%, against 8.22% for Googlebot. ChatGPT spent another 14.36% following redirects, where Googlebot spent 1.49%. If you have reorganized your URLs without redirects, a meaningful share of the crawl budget aimed at you is landing on nothing.

The engine may not fetch you at all. Vercel's team noted that when they asked Claude or ChatGPT for fresh Next.js documentation, they often saw no corresponding request in their server logs, suggesting the models answered from cached or training data. Publishing a page does not guarantee it is read, and a rerun of the same question days later may still return the old explanation.

What does an agent brief actually contain?

Four things, none of which live on a homepage.

Comparison pages with real criteria. Not a competitor smear page. A page that states, in plain retrievable sentences, which situations favor you and which favor the alternative. An agent building a shortlist is performing a comparison; give it one written down.

Pricing logic, not just a price. What changes between tiers, what drives cost, and which situation lands on which tier. An agent screening for budget fit needs the rule, not a "contact sales" button. If pricing is truly bespoke, publish the logic that determines it.

Named integrations and supported standards. Written as entities, not categories. "Connects with your existing stack" is unretrievable. "Snowflake, BigQuery, Postgres, and Redshift" is four entities an agent can match against a constraint.

Outcome-bearing proof with denominators. Not "trusted by leading teams." A named customer, a specific before-and-after, a time window, and a number. This is the social-proof category that the EMNLP work found actually lifts recommendation.

Notice that all four are things a serious human buyer wants too, and that most companies bury below three clicks or behind a form. The agent-readiness problem and the buyer-honesty problem turn out to be the same problem.

How do you run this audit yourself?

Six checks, in order, none requiring a tool.

1. View source on your key pages. Not inspect element, which shows the rendered DOM. View source shows what the crawler gets. If your pricing tiers or integration names are absent from the raw HTML, they do not exist to most engines.

2. Count the specifics. On each core page, count sentences containing a number, named entity, price, or standard. Under five on a page means there is nothing to retrieve.

3. Write down the ten questions a buyer would ask an assistant before they ever hear your name. Constraint-shaped, not brand-shaped.

4. For each question, find the URL on your site that answers it in a self-contained passage. Not "we sort of cover that in the product tour." A specific URL with a specific paragraph. Blanks are your backlog.

5. Crawl your own sitemap for 404s and redirect chains. Given the 34.82% figure, this is higher-return than any copy change.

6. Ask the assistants directly, several times. Same question, multiple runs, multiple engines. Save the raw answers and the citations. Run-to-run variance is large, so a single screenshot proves nothing.

Step 4 is where most teams find the actual work. When I ran it against our own question list, the blanks clustered in one place: every question involving a specific constraint had no home. The gap is rarely a badly written homepage. It is that the comparison page, the pricing-logic page, and the integration page were never built, because no human ever asked for them in a sales call.

What should you build first?

The page that answers the question your best-fit buyer asks before they know you exist. Rank the ten questions from step 3 by how close they sit to a purchase decision, and build downward from the top.

For most B2B companies the order comes out roughly the same. Comparison against the one alternative you actually lose to. Pricing logic. Integration and compatibility specifics. Then the outcome proof that makes the first three believable. The homepage does not appear on this list, and does not need to change at all.

There is a version of this advice that turns into "publish more pages," which is the failure mode this whole category keeps producing. The list above is short on purpose. Four pages that answer real constraints will do more than forty that restate positioning in different words.

What happens when every competitor builds these pages too?

The advantage compresses, and you should plan for that rather than be surprised by it. A 2026 analysis of brand dynamics in LLM recommendation systems measured a substantial first-mover gain that shrank as competitors adopted the same strategy. C-SEO Bench found the same shape from a different direction: as adoption rates rise, marginal gains fall, and the authors describe the dynamic as congested and zero-sum.

The 2024 GEO research showed the same thing in its saturation condition. When every source in the retrieval set was optimized at once, the top-ranked page lost 30.3% of its visibility while the fifth-ranked page gained. Optimization redistributes; it does not manufacture.

So the honest framing is this. Building your agent brief early is worth doing, and the window where it produces outsized returns will close. What does not compress is whether the underlying claims are true, specific, and verifiable, because that is the part competitors cannot copy by running the same playbook. A comparison page is easy to imitate. A comparison page that is accurate about where you lose is not.

How does Trovance help with this?

By starting from the questions rather than the pages. Trovance runs the buyer questions that shape your shortlist across AI systems, preserves the raw answers and citation patterns, and shows which of the four evidence types is missing when you are absent or described wrong.

That ordering matters more than it sounds. The audit above is runnable by hand, and you should run it. What is hard by hand is the diagnosis: knowing whether you were left off a shortlist because the comparison page does not exist, because the pricing logic is unretrievable, because a competitor's third-party coverage is stronger, or because that prompt describes a buyer you should not be chasing. Those four gaps look identical from the outside and lead to completely different work.

Trovance does not guarantee citations, and given how much run-to-run variance sits in these systems, anything that did would be selling you a number rather than a method. What it does is hold the before-state, so when you publish the comparison page and rerun the question a week later, you can tell whether the market's explanation of your company changed or whether you just added a page.

Humans keep the consequential calls throughout: whether a claim is true, whether the proof is sufficient, what can be public, and what deserves to ship.

If you want to see which questions your company is currently missing from, start with your own market questions in Trovance and read the raw answers before you change a single page.

FAQs

Should I rewrite my homepage for AI search?

Usually not. A homepage compresses information to persuade a human who already knows your name, which is a legitimate job. Agents retrieve constraint-specific passages instead. The higher-return work is building comparison, pricing-logic, integration, and proof pages that answer buyer questions your homepage was never meant to address.

What does an AI agent need in order to shortlist my company?

Four things: a comparison page stating which situations favor you, pricing logic explaining what drives cost between tiers, named integrations written as specific entities, and outcome proof with denominators. Each must appear as a self-contained passage in server-rendered HTML at a stable URL.

Can AI crawlers read JavaScript-rendered content?

No. Vercel and MERJ's server-log study found none of the major AI crawlers execute JavaScript, including GPTBot, ClaudeBot, and PerplexityBot. They fetch JS files (11.50% and 23.84% of requests for ChatGPT and Claude) without running them. Gemini and AppleBot are the exceptions.

Does persuasive marketing copy help AI recommend my product?

The research suggests the opposite. An EMNLP 2025 study found scarcity and exclusivity framing reduced how often LLMs recommended a product, while social proof increased it. The 2024 GEO research separately found that authoritative rewriting produced no significant visibility gain over baseline.

Why do AI crawlers hit so many 404 pages on my site?

Vercel and MERJ measured ChatGPT's crawler spending 34.82% of fetches on 404s and Claude's 34.16%, versus 8.22% for Googlebot. The crawlers frequently request outdated asset paths. Maintaining redirects and a current sitemap is higher-return than most copy changes.

How do I test whether AI systems can see my content?

View source rather than inspect element, since view source shows the raw HTML a crawler receives. Confirm your pricing, integrations, and key claims appear there. Then ask several assistants the same buyer question across multiple runs and save the raw answers and citations.

Will building these pages still work once competitors do it?

Returns compress. A 2026 analysis measured first-mover gains eroding as competitors copied the strategy, and C-SEO Bench found marginal gains falling as adoption rises. What resists compression is accuracy: a comparison page honest about where you lose is difficult for a competitor to imitate.

Trovance shows how AI systems explain, compare, and recommend your company, identifies the evidence keeping you off the right shortlists, and turns those gaps into proof-backed content your team can publish. See how AI sees your brand.