A lot of AI search advice boils down to "add more detail". Cover every question, list every service, mention every benefit. The research published in the last two weeks of September says that advice is incomplete. Detail and accuracy are two different things, and the systems that rank and rewrite content do not always tell them apart.
1. Unsupported detail can win AI rankings
In one test, descriptions padded with relevant but unsupported claims ranked higher than honest ones.
The GroundedGEO study (21 September 2026) built 1,950 ranking cases from 50 ecommerce queries. It compared clean descriptions against versions loaded with claims that matched the query but had no supporting evidence.
- On one AI ranker, the unsupported versions gained between 0.065 and 0.092 in normalized rank over the clean versions.
- On a second model the effect was marginal, and on a third it did not appear at all.
- When the researchers gave the system perfect labels saying which claims were supported, the share of unsupported-rich candidates in the top three fell from 65% to 43%.
- But every automatic "evidence judge" they tested failed their pre-set reliability standard against human review.
In plain terms, padding can work on some systems today, and the automatic tools meant to catch it are not reliable yet. That is not a tactic to copy. A roofing page claiming "lifetime warranty on all work" or "certified by every major manufacturer" when that is not true is a risk with customers, and it becomes a liability as checking improves.
Use a claim ledger
Add one table to every content brief and do not publish a claim until its row is complete:
| Exact claim | Supporting passage | Source | Independent? | Applies to (date, location) | Approved? |
|---|---|---|---|---|---|
| "Licensed and insured in Indiana" | License lookup result | State licensing board | Yes | Indiana, checked Oct 2026 | Yes |
| "Most repairs done in one day" | Job records summary | Owner's job log | No | Residential repairs, 2026 | Needs wording |
| "Lifetime warranty" | None | None | n/a | n/a | No, remove |
Caveat: this study used ecommerce queries and open AI ranking models, not Google and not local service results. The researchers say plainly that their automatic defense was not validated.
2. Evidence-based rewrites can break correct copy
Rewriting a page with new sources fixes some errors and introduces others.
The Return or Revise study (24 September 2026) tested 25,870 questions where a system could either keep its first answer or revise it using retrieved evidence.
- Revision repaired 9.32% to 13.42% of answers.
- It also broke 3.05% to 3.33% of answers that were already correct.
- A trained "keep or revise" policy beat always revising by 1.10 to 1.33 accuracy points, yet still let through 38% to 46% of the harmful revisions.
This matters if you refresh old service pages with AI tools, or if anyone on your team rewrites pages whenever a new source turns up. A rewrite can drop an important qualifier ("for asphalt roofs only", "in most of Marion County"), swap an exact figure for a vaguer one, or turn a careful statement into an overconfident one.
The four-check revision rule
Keep the old and new versions side by side, and approve the rewrite only if all four are true:
- Every changed claim is supported.
- The important qualifications are still there.
- Dates, locations and service scope are still accurate.
- The new version is clearly more useful, not just different.
If it fails any check, keep the original. A page that was right does not need to be rewritten to look fresh.
Caveat: the study used short factual answers based on Wikipedia passages and automated correctness judges. Long service pages and human editing may behave differently, but the risk it shows is real.
3. Similar AI answers can cite very different websites
Two AI systems can say much the same thing while sending credit to completely different sites.
The source exposure audit (21 September 2026) collected 1,998 query-platform observations from Google and Baidu AI overviews in English and Chinese.
- Answers to matched questions were fairly similar in meaning, with median similarity scores between 0.701 and 0.813.
- But the websites they cited barely overlapped. Google's Chinese and English overviews shared a host overlap score of just 0.034, and Baidu Chinese versus Google English shared 0.008.
The lesson for a contractor is that "ChatGPT and Google both said roughly the right thing" tells you nothing about whether either one cited you. You have to check the sources each platform shows, one platform at a time.
Track citations per platform
| Prompt | Platform | Market | Answer in brief | Sites cited | Your position | Date checked |
|---|---|---|---|---|---|---|
| Best roof repair company in Carmel, IN | ChatGPT | US, English | Recommends 3 local roofers | Competitor A, a directory, a trade association | Not cited | Oct 27, 2026 |
Record competitors, directories, government sites and trade organizations, not just whether your name appears. Those cited sites tell you where to earn mentions. Our free AI visibility checker gives you the buyer questions to run for your trade and city. For why a single question is not enough, see the multi-turn research in why AI recommends some contractors.
Caveat: the queries came from a research dataset, the Chinese versions were machine-translated, and collection ran from Milan on a logged-in Google account. The study measured which sources were visible, not their quality or rankings.
A workflow that puts evidence first
Put the three findings together and you get one workflow for service pages, location pages and research posts:
- Start from the customer decision. What does the homeowner need to know to call you?
- Build the claim ledger. List every claim and its evidence before writing.
- Draft only what the evidence supports. Leave a claim out rather than stretch it.
- Check revisions against the original. Use the four-check rule above.
- Publish, then measure. Track rankings, AI citations per platform, and image-search visibility with the new multimodal report in Search Console.
The biggest change is mindset: stop treating topical detail as enough. Every claim that matters to a buyer should have evidence behind it, and every rewrite should have to prove it improves on the existing copy. This builds on the source-selection ideas in our semantic SEO research roundup.
What this research does not show
- None of these studies tested Google's local pack or Business Profile rankings.
- GroundedGEO used ecommerce queries and open models, not Google.
- Return or Revise tested short factual answers, not full web pages.
- The source audit measured visibility, not which sources rank or convert.
No new controlled local-pack or Google Business Profile ranking experiment met our standard this week. Use this research to improve how you write, refresh and measure content, not as a list of local ranking factors.
Research sources
- GroundedGEO: Auditing the Evidence Gap in Generative Search Rankings (21 September 2026)
- Return or Revise? Learning When Revision Helps Retrieval-Augmented QA (24 September 2026)
- Auditing Source Exposure in Baidu and Google AI Search (21 September 2026)
Frequently asked questions
Does adding more detail to a page help it rank in AI search?
Sometimes, even when the detail is not supported. In the GroundedGEO study, unsupported but relevant detail lifted rankings on one of three AI rankers tested. That is a reason to be careful, not a tactic: claims you cannot back up are a risk to customers and to trust once systems get better at checking evidence.
What is a claim ledger?
A simple table in your content brief that lists every claim on a page with its supporting passage, its source, whether that source is independent, the date or location it applies to, and whether it is approved. Claims without support do not go live.
Should I let AI rewrite my old service pages?
Only with a review step. In the Return or Revise study, evidence-based rewrites fixed 9 to 13 percent of answers but broke 3 percent of answers that were already correct. Keep the old and new versions side by side and approve a rewrite only when it passes clear checks.
If ChatGPT and Google give similar answers, do they cite the same websites?
Not necessarily. One study found that answers to matched questions were similar in meaning while the websites they cited barely overlapped. Track which sites each AI platform cites, separately, for the questions your customers ask.
Did any of these studies test Google Business Profile or the map pack?
No. They used ecommerce queries, open AI models, short factual questions and a cross-language audit. None tested local pack or Google Business Profile rankings, so use them to improve content quality and measurement, not as local ranking rules.