More homeowners now ask an AI assistant to research contractors for them: "find me a roofer who does storm repairs and works with insurance," or "which electricians near me install EV chargers, and roughly what does it cost?" The assistant searches, opens a handful of pages, pulls out the facts, and writes an answer, sometimes with a recommendation.
That second step, opening your page and pulling out the facts, is where many contractor websites quietly fail. Here is what the new research shows, what it means for a home-service business, and what to check this month.
1. Agent-ready businesses were recommended almost twice as often
In the largest of the four studies, businesses whose websites were easy for AI agents to read were clearly recommended 1.9 times as often as comparable businesses whose sites were not.
The paper, AX Is the New AEO (published 28 September 2026), ran 37,927 "agent journeys" (an AI assistant answering a buyer's question about a specific business) across 1,056 real businesses and four different AI setups. The researchers matched businesses on how well known they were and how much the AI already knew about them, then compared sites that were easy for an agent to read with sites that were not.
The results were consistent across all four setups:
| What was measured | Agent-ready sites | Other sites |
|---|---|---|
| Answers built from the business's own pages | about 78% | about 56% |
| Clear recommendation of the business | about 20% | about 11% |
Two more findings matter for contractors. First, only 7–10% of the finished answer came from what the AI already "knew" from training. Almost everything came from what it could fetch at the time. Second, the main failure was not the AI inventing facts. It was leaving them out: when the answer had to be assembled from other websites, it was 3.7 times more likely to contain none of the facts the buyer actually asked for.
For a contractor, that means if the AI cannot read your pricing, service area, or process from your own site, it will fill the gap from directories, aggregators, review sites, or a competitor's page. Or it will simply give the homeowner a vaguer answer that does not mention you.
One honest caveat: this was an observational comparison, not a controlled before-and-after test. The researchers used their own readiness score and did not match sites on how much content they had. Treat it as strong directional evidence, not proof that fixing your site will double recommendations.
What "agent-ready" means for a contractor website
An agent-ready page is one where the important facts are present in the page's text when it is fetched, without needing JavaScript, clicks, or images to reveal them.
Most of the problems are invisible to a human visitor. The page looks fine in a browser, but the facts an AI needs are hidden:
- Page builders that render everything with JavaScript. Some site builders send an almost empty page and fill it in with scripts. Many AI fetchers do not run those scripts.
- Prices and service areas inside images. A graphic that says "Serving these 12 towns" or a pricing table saved as a picture is unreadable text.
- Key details behind tabs, accordions, or "read more" buttons that only load content after a click.
- Quote calculators and booking widgets that hold the only mention of price ranges or availability.
- Aggressive bot blocking from security plugins or firewall settings that turns AI crawlers away entirely.
Run the plain-text fetch test
Open your most important service page and view it with JavaScript turned off, or use your browser's "view source" and search for your key facts. If you cannot find them in the raw page, an AI agent probably cannot either. Check that each important service and location page contains, in plain text:
- Your business name and the exact service name
- The service area (the towns or region you actually cover)
- Pricing, a price range, or how quotes are worked out
- Your process and typical job duration
- Who the service is for, and what you do not do
- How to contact or book you
If a fact is missing from the plain text, add it as a sentence or a simple HTML table. That is the single highest-value change from this week's research.
2. Mentioning your service is not the same as answering the question
A page can be clearly about the right service and still not contain enough information to answer the buyer's actual question.
The second paper, Relevance Is Not Sufficient Evidence (27 September 2026), tested twelve AI models on questions where the retrieved pages were relevant but did not actually contain the answer. Even when told to say "I don't know," the models answered anyway between 40% and 99.3% of the time. The researchers' proposed method, which checks whether the evidence actually covers the question before answering, scored 0.837 against 0.746 for the best earlier approach.
For contractors, the practical lesson is simple. A roofing page that says "we offer roof replacement" is relevant, but it cannot support an answer to "how much does a roof replacement cost?" or "do you handle insurance claims?" When your page is relevant but silent, the AI either guesses or borrows the answer from someone else's page.
Before you publish or update a service page, run a quick evidence check:
| Question a homeowner will ask | Does the page answer it directly? | What qualification belongs next to it? |
|---|---|---|
| What does it cost? | Yes / No | Size, materials, location, date of the estimate |
| Do I qualify, or is this for me? | Yes / No | Property types, exclusions, minimum job size |
| How long does it take? | Yes / No | Weather, permits, parts availability |
| Is there a warranty? | Yes / No | What it covers and for how long |
| Are you licensed and insured? | Yes / No | Which licences, which areas |
If a row is a "No," either add a direct answer or state plainly that it depends and on what. A range with honest conditions is far more useful to an AI, and to a homeowner, than silence.
Caveat: this study tested question-answering datasets, not search rankings or full web pages, so it tells us how AI models behave with weak evidence rather than how any search engine ranks a page.
3. Keep your text, tables, and photos saying the same thing
Facts are easier for AI systems to connect when your paragraphs, tables, photo captions, and diagrams use the same names for the same things.
The third paper, PILAR (26 September 2026), linked facts from sentences, tables, and figures through shared names while keeping track of which page each fact came from. It improved accuracy modestly: 1.6 points overall and 5.9 points on harder three-step questions. Most of the gain came from the text, not the images.
The contractor takeaway is about consistency. If your copy says "asphalt shingle replacement," your pricing table says "shingle roofs," and your project photo is captioned "recent project," those three pieces do not obviously connect. A caption like "Asphalt shingle roof replacement after hail damage, completed in four days" connects the photo to the service, the problem, and the timeline you described in the text.
Use the same terms for the service, material, problem, and location across your headings, paragraphs, tables, image alt text, and captions.
Caveat: this was a document question-answering experiment, not an SEO study. The gains were small and not consistent across every AI setup.
4. Do not judge your AI visibility from one question
Asking an AI assistant one question about your trade shows you very little about whether you get recommended during a real decision.
The fourth paper, Conversational Capture (30 September 2026, to be presented at HAI '26), argues that a source cited early in a conversation becomes more likely to be cited again, both because the AI leans on earlier sources and because the user's follow-up questions build on them. In the authors' model, rankings measured from a single question and rankings measured across a whole conversation agreed only weakly (a correlation of 0.4).
Real homeowners do not ask one question. They ask what a job costs, whether it suits their home, whether you cover their area, how you compare, and then who to hire. A better test looks like this:
- A broad service question ("What's involved in replacing a water heater?")
- A cost or suitability follow-up
- A location or eligibility question
- A comparison request
- A request for a recommended provider
Record which websites are cited and whether your business is mentioned at each step, then repeat the same questions in a fresh conversation to see how much the earlier answers influenced the later ones.
Caveat: this paper is a theoretical framework with a modelled example. It does not prove the effect happens in ChatGPT, Gemini, or Google's AI features today, so use it as a better way to test, not as a confirmed ranking factor.
What to do this month, in priority order
Fix whether AI can read your facts before you spend any time polishing copy for AI.
- Run the plain-text fetch test on your homepage and your top five service pages. Move any key fact that is locked in images, scripts, or click-to-reveal elements into plain text.
- Run the evidence check on each of those pages, and add direct, qualified answers to the cost, eligibility, timing, warranty, and licensing questions.
- Align your wording across headings, tables, captions, and alt text so the same service and problem are described the same way everywhere.
- Test visibility as a conversation, not a single question, and log which sites get cited at each step.
None of this replaces the fundamentals. Your Google Business Profile, reviews, and service pages still decide whether you are found at all. Our guide to how contractors get found in AI search covers those basics, and our earlier research roundup explains why verifiable business information matters to AI systems.
What this research does not show
None of the four studies tested Google's map pack or Google Business Profile rankings, and none ran a controlled experiment on live contractor websites.
The strongest finding (agent-ready sites recommended about twice as often) is observational. The other three papers test AI behaviour in lab settings or in theory. No new controlled study on local-pack or Business Profile rankings met our standard this week. The direction is still consistent across all four: AI systems reward first-party facts they can read and verify, and they fill gaps with weaker sources when they cannot.
If you want to know whether AI assistants can read your service pages, and what they say about your business today, our AI visibility service covers exactly that, or you can start with a free SEO audit.
Frequently asked questions
What is an agent-ready website?
An agent-ready website is one where an AI assistant can fetch a page and read its important facts in plain text, without running JavaScript, clicking tabs, or reading images. For a contractor, that means the service name, service area, pricing approach, process, eligibility, and contact details are all present in the page text.
Do AI assistants really recommend agent-ready businesses more often?
In a September 2026 study of 37,927 AI research sessions across 1,056 businesses, agent-ready sites were clearly recommended about 1.9 times as often as comparable sites. The study was observational rather than a controlled experiment, so it shows a strong association rather than guaranteed cause and effect.
How can I tell if AI can read my contractor website?
View your key service pages with JavaScript turned off, or open the page source and search for your prices, service area, and process. If those facts only appear in images, scripts, booking widgets, or behind clicks, many AI fetchers will not see them.
Should I publish my prices for AI search?
You do not need exact quotes, but giving a realistic price range with the factors that change it gives AI assistants something accurate to work with. When a page has no pricing information at all, an assistant may guess or pull a figure from a directory or a competitor instead.
Does this affect my Google map pack rankings?
None of these studies tested the Google map pack or Business Profile rankings directly. They concern how AI assistants research and recommend businesses. Map pack visibility still depends mainly on relevance, distance, prominence, and reviews.
How should I test whether AI recommends my business?
Test a full conversation rather than one question: ask about the service, then cost, then eligibility or location, then a comparison, then a recommendation. Note which sites are cited at each step, then repeat the conversation in a fresh session.