a16z's newest cohort is harder for AI to find than YC's. We scanned both.

A quick recap for anyone who missed part one. When a buyer asks ChatGPT or Claude for the best tool for their problem, the model builds a shortlist. To be on it, your site has to clear two very different bars. First, an AI has to be able to reach your content, actually fetch and read it. Second, it has to be able to tell what you are, sort you into the right category, understand what you sell, quote you correctly. The first is a plumbing problem. The second is a labeling problem. In the YC Spring 2026 batch, which we scanned last month, reach was mostly fine and labeling was mostly missing.

So we wanted to know: is that a YC quirk, or a pattern? a16z's Speedrun is a different program with a different selection filter, and SR006 is its newest and largest cohort, a revenue-first group where the bar for admission is a working product with paying customers. If any group of founders were going to care about being found, it would be this one. We scanned all 59 at the exact website each lists in a16z's own directory, with the identical engine and checks we ran on YC.

Same shape. Reach is strong in both. Labeling collapses in both. And the revenue-first cohort labels less, not more.

One picture tells most of the story

Here are both cohorts on the same eight checks. The further a point sits from the center, the more of that cohort passes it. The two shapes are nearly identical at the top, where reach lives, and both pinch inward at the bottom, where labeling lives. The blue SR006 shape pinches further in on the structured-data axes.

What is actually different, and what only looks different

At 59 companies, a gap of a few points is two or three sites, which is noise, not signal. So we ran the numbers properly. Of the sixteen checks we compared, only five differences hold up as statistically real. The rest are ties dressed up as differences. Here are the five that survive.

Every gap in that table clears an exact statistical test. Everything not listed, the empty-shell rate, content served, on-page FAQ, canonical, HTTPS, crawler access, all the AI-bot blocking, Core Web Vitals, and the median words a crawler sees, came out a statistical tie between the two cohorts.

Read the table and a clear pattern falls out. On four of the five real differences, SR006 is weaker, and all four of those are about the same thing: labeling. Less schema, less usable schema, far less FAQ markup, less image alt text. The one check SR006 wins is the sitemap, which is plumbing, the mechanical file that lists your pages for a crawler.

SR006 built the roads and skipped the signs. Strong on the mechanical layer, thinner on the layer that tells a machine what any of it means.

The pattern is bigger than either cohort

This is the part that matters more than the head-to-head. Two independent groups, chosen by two different programs with two different filters, months apart, produced the same shape: highly reachable, poorly labeled. That is what turns a single scan into a finding. If only YC looked like this, you could call it a fluke of one batch. Two elite cohorts pointing the same way is a pattern.

And the direction of the difference is the counterintuitive part. SR006 is the revenue-first cohort, companies with paying customers before demo day. The intuition says that group should be sharper about being discovered, because their revenue depends on it. The data says the opposite: they are the ones who label less. The most plausible read, and it is a read, not a proof, is that a team heads-down on shipping product and closing revenue treats the marketing site as an afterthought, exactly the pages an AI reads first. Being good at revenue does not make you legible to the machine that increasingly decides who gets recommended. If anything it competes for the same attention.

We are careful not to overclaim the why. What the numbers support is the what: across two elite cohorts, the semantic-labeling layer is the consistent weak point, and it is weakest in the cohort you would least expect.

The five that scored highest in SR006

We name the best, not the worst. But we are going to be precise about what best means here, because it makes the cohort's gap visible even at the top. These five posted the strongest results on our measured checks, at the exact URL shown (each verified to score the same on its apex and www forms, so there is nothing to explain away). What they have in common is reach. What most of them still miss is full labeling.

Look at the right-hand column. Only two of the five, Antihero Studios and Mirror Mirror AI, actually clear both bars: reachable content and real machine-readable labels including FAQ markup. Clair Health and Grove Tax serve content and carry some schema but skip FAQ markup. Amdahl, one of the highest scorers on reach and the technical basics, carries no schema type a machine recognizes at all. When the labeling gap shows up even in your top five, it is not an accident at the bottom of the distribution. It is the cohort.

All five, like almost every company in both cohorts, fail the strict Core Web Vitals bar. More on that in the method note.

The fix list is the same as last time, because the gap is the same

If you are in SR006, or building anything like it, the shortlist has not changed. Add Organization plus your real product type (SoftwareApplication, Product, or Service) schema, the single biggest gap in this cohort. Add a genuine FAQ with FAQPage markup, which only 1 in 20 of SR006 has. Give your images alt text. You already ship a sitemap, that part you have handled, so this is a couple of hours of labeling on top of plumbing you have already done. And if you were told to add an llms.txt for Google, skip it for that purpose: Google has said on the record its Search systems do not use the file, and named it directly in its 2026 AI guidance as a tactic that does not help. Some coding agents and assistants do read it, so it is cheap agent-facing infrastructure, just not a Google lever.

How we did this, and what to distrust

Read this before you argue with the numbers, because it is where the honesty lives.

The data. Every figure comes from one source, our live scanning engine, the same one anyone can run on their own site. No number was estimated, modeled, or written by an AI. The engine fetches each site, renders JavaScript-heavy ones with a real headless browser, and reports deterministic facts: how many words a crawler sees, whether robots.txt blocks a given bot, which schema types are present.

The two cohorts. YC Spring 2026 (197 companies, 195 evaluated) scanned 2026-06-22, and a16z Speedrun SR006 (59 companies, all 59 evaluated) scanned 2026-06-29. SR006 was enumerated first-party from a16z's own company API by its cohort field, and each company was scanned at the exact website it lists there. Identical engine, identical checks, identical thresholds, so the comparison is like for like.

Small samples, stated plainly. At 59 companies, a gap of a few points is only two or three sites and is not meaningful. We tested every difference for significance and report only the five that hold up under an exact test. We say so explicitly wherever a gap is a tie rather than a difference, and we do not build any claim on the ties.

What we do not claim. The scores behind these signals are page-level and specific to the URL scanned; we compare measured yes or no signals, not a single composite number. Core Web Vitals here are lab measurements on a cold load, which run pessimistic against real users and are a low-weight check, so read the near-universal CWV failure as almost nobody clears the strict lab bar, not as these sites feel slow. And the reading of why revenue-first founders label less is a hypothesis, not a measured fact. The measured fact is the gap itself.

Want to know which side of this your own site is on, reachable but unlabeled, or both? Scan it at potatometer.com. It takes about thirty seconds and shows you exactly what a crawler sees before any JavaScript runs.

← Back to Blog