In October 2025, an analyst named Jes Scholz ran 100 conversations through ChatGPT Agent mode and watched what the agent actually did when it landed on a webpage. The result rewrote what we thought we knew about technical SEO. Sixty-three percent of the time, the agent took one look at the page and walked away.
If you run a small business website in the UK, that figure should stop you cold. It means that even when an AI agent is sent to your site by a real customer query — "best Spanish restaurant in Tunbridge Wells", "luxury holiday let in Cornwall with hot tub", "barber in Newark-on-Trent open Sunday" — there is roughly a two-in-three chance that the agent never reads a single line of your content before going elsewhere.
The reasons are technical, fixable, and almost entirely invisible to Google Analytics. This post breaks down the five most common causes, and what to do about each one this week.
The mistake most small business owners make is assuming an AI agent behaves like a person with a browser. It does not. When ChatGPT Agent lands on your homepage, it is not appreciating your hero image, scrolling through your testimonials, or pausing on your team photos. It is parsing raw HTML in a few hundred milliseconds and asking one question: can I extract a structured answer from this page right now?
If the answer is no, the agent leaves. There is no second chance, no remarketing, no abandoned-cart recovery. The agent moves to the next URL in its candidate list and the customer never knows your business existed.
Scholz's research identified five specific technical reasons agents bounce. They appear in roughly this order of frequency on small business sites I audit at MapRank.
Forty-six percent of ChatGPT visits begin in reading mode: a text-based browser with no JavaScript execution, no CSS, no images, no schema markup rendered client-side. The agent sees the raw HTML response from your server and nothing else.
For a site built on a typical no-code platform with full server-side rendering — WordPress, Shopify, Squarespace, Hostinger Website Builder — this is fine. The HTML is complete on first load. But for sites built as JavaScript single-page applications without server-side rendering (a common pattern with React, Vue, or some headless setups), reading mode shows the agent something like this:
<!DOCTYPE html>
<html>
<head><title>Loading...</title></head>
<body>
<div id="root"></div>
<script src="/static/js/main.js"></script>
</body>
</html>To a human in Chrome, that page is a fully interactive application. To a ChatGPT agent in reading mode, it is an empty room. The agent leaves.
This is exactly what we found in the Newark Traditional Barbers audit: a JavaScript SPA with near-zero crawlable HTML. Strong reviews, good local rankings on ThreeBestRated, 293 customer ratings — and yet effectively invisible to AI retrieval bots because the homepage rendered as an empty shell on first request.
curl -A "ChatGPT-User" https://yourdomain.com from a terminal. If the response does not contain your headline, your service descriptions, or your contact details in plain text, you have a reading-mode problem.Agents are intolerant of friction. A 404 (not found) or 500 (server error) on a page they were directed to ends the conversation instantly. Worse, a 301 redirect to an unexpected URL — for example, your old blog post redirects to your homepage rather than the updated article — confuses the agent's parser and often triggers an abandonment.
This matters more than it used to because of how ChatGPT discovers content. The agent typically calls the Bing Search API, gets back a list of candidate URLs, and tries them in order. If your URL is on that list but returns an error or a confusing redirect, the agent skips you and tries the next candidate. You lost your spot before anyone knew you were in the running.
/old-page redirects to /new-page which then redirects to /newer-page, the agent often gives up halfway. Make redirects single-hop.This is where the technical picture changes most sharply between human SEO and GEO. For a human visitor, Google has tolerated First Contentful Paint times up to about 1.8 seconds without significant ranking penalty. For ChatGPT agents, the bar is dramatically tighter.
Recent analysis by Superlines found that pages with First Contentful Paint under 0.4 seconds were three times more likely to be cited by ChatGPT's research agents than slower competitors. Four hundred milliseconds. That is roughly the time it takes to blink.
Why so strict? Because agents are processing many candidate URLs in parallel against a budget. If your page takes two seconds to deliver its first byte of meaningful content, the agent has already moved on to the next candidate that delivered in 200 milliseconds.
defer or async.This one is delicate. Cloudflare changed its default configuration in late 2024 to challenge or block AI crawlers more aggressively. Many small business owners have these defaults active without realising it. The result: ChatGPT agents arrive at the page, hit a Cloudflare challenge, and leave because they cannot solve the CAPTCHA.
The robots.txt file is the other common offender. A surprising number of sites still carry blanket disallow rules from old SEO guides that were written before AI agents existed. The most common pattern looks like this:
User-agent: *
Disallow: /That blocks everything, including ChatGPT-User, OAI-SearchBot, PerplexityBot, and Google's Gemini crawlers. If you ever wondered why your competitor appears in AI answers and you do not, this is often the reason.
This is where small business owners need to be precise. There are three categories of OpenAI bots, and they do different things:
For most UK small businesses, the right answer is to allow all three. The opportunity from being cited and visited by AI agents far outweighs any concern about training data — particularly for service businesses where the content is local, contextual, and regularly updated.
https://yourdomain.com/robots.txt in a browser. If you see any Disallow: lines that block the AI user-agents above, remove them.Even when the agent successfully loads your page, it can still leave for a more subtle reason: it cannot find the answer to the user's question quickly enough in your content structure.
A landmark February 2026 study by Kevin Indig analysed 1.2 million ChatGPT responses and found that 44.2% of all citations come from the first 30% of webpage content. The pattern is so strong that Indig calls it a "ski ramp" distribution. If your direct answer is buried in paragraph 14 after a thousand-word introduction, the agent will not find it before deciding the page does not match the query.
<h1>, <h2>, and <h3> hierarchy helps agents map the content. One H1 per page, descending logically.curl -A "ChatGPT-User" against your homepage and key service pages. Confirm full HTML content is returned.The hardest part of all this for a small business owner is that none of it shows up in your analytics dashboard. ChatGPT agents typically use text-based browsers that do not execute the JavaScript tracking code in Google Analytics 4 or Plausible. They appear as either nothing at all, or as anomalous "direct" traffic with zero engagement metrics.
The only ground-truth source is your server access logs. Hostinger, cPanel, and most managed hosts provide raw log access. The user-agent strings to look for are:
ChatGPT-User # Agent fetches (real-time, user-triggered)
OAI-SearchBot # Live search crawler
GPTBot # Training crawler
PerplexityBot # Perplexity equivalent
ClaudeBot # Anthropic crawler
Google-Extended # Gemini trainingA 30-day site log study by Digital Applied across 12 production websites found a median of 690 user-triggered AI agent fetches per site per day. That is invisible traffic — real customer queries reaching your server, parsed by agents — and the only way to see it is in the logs.
For most MapRank clients, my first deliverable on a GEO audit is a 30-day log analysis showing exactly which AI bots are reaching the site, which pages they fetch, and what HTTP status codes they receive. It is the closest thing we have to AI Search Console.
Before any of the technical fixes above matter, you need a baseline: how visible is your business currently in AI-driven discovery, and how do you compare to the businesses you are competing with for the same queries?
To answer that without a full audit, I recently built and deployed the Share of Model Analyzer — a free tool you can run on your own site in about 90 seconds.
You enter your URL plus up to three competitors. The tool returns an Entity Score (0-100) for each, a side-by-side comparison of your structural readiness signals (schema coverage, NAP consistency, authority citations, content depth, technical accessibility), and the three highest-leverage quick wins specific to your business and sector.
It is the same first-step methodology I use on every MapRank engagement, packaged into a tool you can run yourself. You get the report by email and WhatsApp.
Run the Share of Model Analyzer →
If your Entity Score comes back below 30 and your competitors are scoring 50+, the technical issues in this article are almost certainly the reason. The good news is that the seven-day fix list above closes most of that gap.
If you want the deeper analysis — server log review, robots.txt audit, FCP and TTFB measurement, and a prioritised fix list — I run free preliminary audits for UK and EU small businesses. Turnaround within 72 hours.
Request your audit →