The AI Reputation Gap: What a Model Says About You

🇹🇷 Türkçe: Bu yazının Türkçesini oku →

A client of 97th Floor sells venue management software. They are one of the larger players, and their growth plan was to move down-market and win the smaller, newer venues. Sensible strategy, decent product, real budget behind it.

Then somebody asked an AI assistant what it thought of them. It came back with a description their own marketing team would never have written: the IBM of venue management software. Well established, yes. Also clunky, hard to use, and not for small venues.

Nobody at that company had approved that sentence. Nobody had written it. It was simply what the machine had concluded from everything of theirs it could reach, and it was being handed to the exact buyers the growth plan depended on.

I have been watching this pattern arrive at client after client this year, and the reason it keeps catching people is that it is invisible from the inside. Your website looks right. Your messaging is current. You have no way of knowing that the answer being given about you is a different answer, unless you go and ask.

The gap is expensive because the research now happens before you are in the room

Paxton Gray, CEO of 97th Floor, told that story in the first session of an AI discoverability series he runs with Jodi Cerretani, CMO of WordPress VIP. He has been doing search work for two decades and does not reach for superlatives, so this one is worth quoting exactly:

we’ve been doing SEO for a little over 20 years now, and I have to say there’s not been a bigger update than what’s happened with AI search. We went from 10 blue links and trying to get shown there, and now what we need to do is get recommended.

Paxton Gray, CEO, 97th Floor

Shown versus recommended is the whole shift in four words. A ranking is a position you occupy in a list the buyer then works through. A recommendation is a sentence someone reads and acts on, and you are not in the room while it is being made.

The numbers behind that are firmer than they were a year ago. G2’s 2026 buyer research, a March 2026 survey of 1,076 B2B software buyers and decision makers, reports that:

• 51% now start their research with an AI chatbot more often than with Google, up from 29% in the previous year’s report.
• 69% chose a different vendor than they had originally planned, on that guidance. One in three bought from a vendor they had never heard of.
• 8 in 10 say it made the decision happen faster.

Read the second one again. It is not a visibility statistic, it is a switching statistic. Two thirds of buyers ended up somewhere other than where they set out for, and the thing that moved them was a paragraph of generated text about a set of vendors. If that paragraph describes you as the clunky incumbent, you are losing deals in a conversation you never joined.

The model is not reading your homepage. It is reading everything.

The instinct when this lands is to go and fix the messaging on the site. That is where most of the budget goes and it is mostly wasted, because the site is only one of the inputs.

Jodi Cerretani calls the real input the total information environment: the live site, and also “that micro site that your business unit launched in 2019 and is still up, or a legacy subdomain that nobody sunset, or an old product PDF that’s no longer up to date, press coverage, Reddit threads.” All of it goes into the same pot.

And then the part that makes it a commercial problem rather than a tidiness problem:

when AI is consuming all this information, it’s normalizing or averaging what it sees, and an average brand is a diluted brand.

Jodi Cerretani, CMO, WordPress VIP

She reports comparing six sites belonging to one company and finding comprehension ranging from 88% at the high end down to 59%. Same brand, same products, and a four-in-ten chance the model is working from the weak version.

This is the bit I would push hardest with anyone about to sign off an AI visibility programme. Averaging means your worst property is not neutral. It is actively pulling the answer down, and a microsite from 2019 with three pages of superseded positioning gets a vote alongside the work you shipped last quarter.

Being on the site is not the same as being readable

The second failure mode is quieter and I see it more often. Cerretani’s mental model for it is the one I have started borrowing: an AI crawler is “a brilliantly fast reader, but it’s incredibly literal.” It does not care about your design. It reads structure.

In one assessment she describes an enterprise brand whose sixteen product pages carried no markup at all. To a person each page looked fine, handsome even. To a model they were very difficult to decipher. In another, product pages were missing from the sitemap and their content sat inside JavaScript, which renders empty boxes to most crawlers. One set of those pages took 233 visits from human beings and six from AI crawlers. Not six percent. Six.

Your content can be live, it can be on brand, it can be perfectly written, it can be beautiful. But it can be functionally invisible to the systems that your buyers are now engaging with first.

Jodi Cerretani, CMO, WordPress VIP

Worth knowing that this is not only a small-company problem. Cerretani says VIP ran the same audit against their own site and found their server layer rejecting one of Anthropic’s crawlers. Elsewhere, a regulatory compliance plugin at a global enterprise was turning away around 15% of ChatGPT requests, roughly 68,000 of them in the measured window, because it was treating AI crawlers as human visitors from blocked countries.

Nobody chose any of that. It is what a stack does when no one has been asked to check.

Inside a real WordPress or WooCommerce build, this is four jobs

Session one closes on five asks to take back to a team. That is the right shape for a webinar and too abstract to price, so here is what each one is when it reaches us as a ticket, with rough effort for a mid-market estate.

• The property inventory, and it starts with hosting rather than marketing. Ask whoever holds the hosting accounts for every domain and subdomain, not the marketing team for every site they run. The difference between those two lists is the finding, every time. Then decide, per property, whether it is switched off, redirected or brought current. Half a day to produce the list, and the decisions are a meeting, not a project.
• A structural pass at template level, not page level. One H1 per page, a clean heading hierarchy, and schema on every template that matters: product, article, organisation. On WordPress this is theme work, so the cost scales with how many templates you have rather than how many pages. Two to four weeks for a typical commerce theme, and it doubles as an accessibility win, which is usually what gets it funded.
• Sitemap against reality, plus a text fallback behind every JavaScript component. Diff the sitemap against the pages you actually want found. Then request your own key pages with something that runs no JavaScript and read what comes back. If a product page returns an empty shell, that is your 233-versus-six problem and it is a rendering decision, not a content one.
• Crawler traffic out of the server logs. The cheapest of the four and the one nobody runs, because JavaScript analytics cannot see it. Grep the access logs for AI user agents and put the table in front of whoever owns the edge configuration. A morning’s work, and it is the only place a wrongly configured plugin or firewall rule shows up.

The fifth ask is to name an owner, which is a different kind of problem and gets its own piece later in this series.

One honest caveat on the server logs, because this is where confident write-ups tend to overreach. You cannot settle crawler access from outside. Edge providers admit verified crawlers by network identity rather than by what a request calls itself, so sending traffic with a crawler’s user agent proves nothing about whether that crawler is allowed in. The logs are the only source of truth, which is exactly why the ask is worded as show me, not check whether.

What to do with this on Monday

Before any of the four jobs above, spend twenty minutes establishing whether you have a problem at all. Ask three or four assistants what your company does, who it is for, and how it compares to the two competitors you lose to most often. Do it without logging in, so you get the generic answer rather than one shaped by your own history.

Then read the answers the way a buyer would. Not “is this flattering” but “would this shortlist us.” If what comes back is a version of the company you were two products ago, or a description that is technically accurate and commercially useless, you have found your baseline. That is the whole point of the exercise: Gray’s argument is that you cannot fix the gap until you can see it, and the seeing is free.

What it is not is a content problem, which is where most teams reach first. It is a properties problem, a structure problem and an access problem, in that order, and the content strategy only starts paying once those three are answered.

This is part one of a series working through what Cerretani and Gray put on the record across their three sessions, with the implementation detail they leave out. Next: why being readable and being recommended are two different jobs. If you want the four jobs scoped against your estate, come and talk to us.