Downer Brothers made one list out of four
It had the deepest review history any engine reported near this town, yet three engines left it out.
Field Study 001 · Local AI Search
One local search question. Eleven different answers.
We asked ChatGPT, Perplexity, Microsoft Copilot, and Claude to name the best landscaping service in Andover, Massachusetts. Only two businesses appeared on every list.
4AI assistants
1identical query
12weeks between tests
0star-rating changes
Method
Five words, one ordinary buying decision
“Best landscaping service in Andover.” A homeowner types it, an engine answers, and somebody’s truck pulls into the driveway on Saturday.
We captured the full response from four assistants on May 8, 2026, then ran the same question again on August 2. We did not contact any of the businesses, buy anything, or steer the engines toward a preferred answer. Between the two runs, we changed nothing.
Andover proved to be a useful test market: a review economy measured in dozens rather than thousands, a few nearby towns in the mix, and seven or eight serious landscaping firms. Less than three months separated the snapshots. Plenty moved.
The evidence
The answer matrix
On August 2, four engines named eleven companies. T&B Landscape & Irrigation and Andover Landscape Design & Construction were the only two every engine surfaced.
| Business | Copilot | ChatGPT | Perplexity | Claude |
|---|---|---|---|---|
| T&B Landscape & Irrigation | 3 | 1 | 2 | 6 |
| Andover Landscape Design | 4 | 2 | 1 | 1 |
| Pelletier Landscaping | 2 | — | 3 | 5 |
| Alvarez Landscaping | 1 | 4 | — | — |
| Ground Care Landscaping | 5 | — | — | 4 |
| C&C Landscaping Maintenance | 6 | — | — | 3 |
| Downer Brothers Landscaping | — | — | — | 2 |
| Jireh Landscaping & Irrigation | — | — | 4 | — |
| Northeast Landscape Contractors | — | 3 | — | — |
| DiCicco Landscape & Irrigation | — | — | — | 7 |
| M J Colombo Landscaping | — | — | 5 | — |
Named first Position in responseClaude declined to crown one winner; the number reflects list order.
Alvarez ranked first on Copilot and disappeared elsewhere
ChatGPT gave it a passing mention. Perplexity and Claude did not surface it at all.
Neither business had a bad month. Each had a retrieval problem on some engines and a strong result on another, at the same time.
01
Finding
The contest is over the shortlist, not the ranking
Copilot and Claude reported identical review data in August. Ground Care at 4.8 with 47 reviews. C&C at 4.9 with 30. Pelletier at 5.0 with 14. T&B at 4.7 with 74. Andover Landscape at 4.8 with 29. Digit for digit, the same underlying numbers.
Then Copilot led with Alvarez and never mentioned Downer Brothers or DiCicco. Claude led with Andover Landscape and never mentioned Alvarez. Two engines, one dataset, two different rosters.
The variable deciding who got recommended was not star rating or review volume. What differed was which businesses entered the running before scoring happened.
Most local search reporting still measures position. Position is the last step, and by then the interesting decision has already been made upstream.
A 4.9 does nothing for you inside a candidate set you never joined.
02
Finding
Rankings moved while the businesses stood still
Across the seven businesses tracked on both dates, not one star rating changed. Review counts moved by twelve in total.
Alvarez Landscaping135 → 138
T&B Landscape70 → 74
Ground Care44 → 47
Jireh Landscaping33 → 34
Andover Landscape28 → 29
C&C Landscaping30 → 30
Pelletier Landscaping14 → 14
In May, ChatGPT opened with Downer Brothers as its top pick. By August, Downer Brothers had vanished and T&B held the top spot. Perplexity opened May with Jireh at 4.9. By August, Jireh had slid to fourth.
Copilot did not budge: Alvarez first in May, Alvarez first in August, with the same shortlist in a slightly shuffled order.
The engine leaning on structured listing data held steady. The two engines assembling answers from the open web reshuffled winners while the businesses underneath them did close to nothing.
A recommendation drop is not automatic evidence of a reputation problem. It can be a sourcing problem, and the repair work looks entirely different.
03
Finding
Same company, same week, opposite verdicts
T&B Landscape & Irrigation received two reviews of its standing in early August. They do not resemble each other.
Perplexity · August 2
Safest all-around choice
Perplexity put T&B ahead for lawn maintenance, irrigation, and snow removal. Nearly every claim, including the competitor table, cited one page on T&B’s own website.
Source diet: one self-published pageClaude · August 2
Sixth of seven
Claude put T&B last on rating at 4.7 and filed it under irrigation rather than general contracting. It treated the supply yard as a convenience, not a reason to hire them.
Source diet: aggregated review and profile dataBoth answers hold up inside their own evidence. They were reading different documents, and that alone was enough to flip the verdict.
04
Finding
A vendor page became the reference source
T&B published a page answering the exact question a customer would type, in its own market, naming its own competitors. Perplexity found it and built an answer on top of it.
The page was already among Perplexity’s sources in May. By August, close to the whole answer rested on it. Yet the same page did nothing on Copilot, which ranked T&B third behind a business Perplexity never named.
Publish the definitive page and you move one kind of engine. The other kind may never see it.
05
Finding
ChatGPT changed what it accepts as proof
In May, ChatGPT cited company websites and nothing else. By August, the same query returned BBB accreditation, HomeAdvisor and Houzz ratings, a founding year, and regional design awards.
Its evidence diet shifted from vendor self-description to outside validation. Businesses without maintained third-party profiles did not make the cut, and Google star ratings did not rescue them.
06
Finding
The engines disagree about the facts too
Claude described Downer Brothers as holding the area’s highest review volume at 114. Copilot listed Alvarez at 138 on the same date. Both cannot be right unless the engines draw the map differently, and they do. Claude separated North Andover from Andover. Copilot did not.
Copilot also listed Alvarez with a 978 phone number in May and a 603 number in August. That could reflect a business change, a directory inconsistency, or something worse. We did not verify which.
What is clear: whatever sits in a listing record now flows directly into the sentence a homeowner reads before deciding who to call. Directory errors used to cause a misdial. They now propagate into recommendations.
What we take from it
Six changes to how local visibility gets measured
Review quality still matters. This study shows that it is one input among several, weighted differently by each engine.
Track presence before position
Absence is the loudest result in this dataset. Five of eleven businesses appeared on exactly one engine. A ranking report cannot show the three engines where you are missing entirely.
Audit each engine’s source diet
Copilot answered from listing data, ChatGPT from third-party validation, Perplexity from a published page, and Claude from aggregated profiles. Four pipelines produced four different answers.
Establish a repeated baseline
Two dates show movement, but not whether it came from a lasting change or run-to-run variance. Serious measurement needs a consistent sampling cadence.
Treat directory hygiene as answer hygiene
The Alvarez phone-number discrepancy flowed directly into a recommendation. Listing errors no longer stop at a directory. They become sentences customers act on.
Build the third-party footprint deliberately
BBB, Houzz, HomeAdvisor, and a regional award helped decide ChatGPT’s August shortlist. On that engine, those profiles were the basis for selection.
Publish the definitive page, with limits
T&B’s comparison page carried Perplexity almost single-handedly and did nothing for them elsewhere. It is worth doing, but it is not an across-the-board fix.
Method and limits
What this study cannot tell you
Two dates in one small town will only carry an argument so far.
- One market, one query phrasing, and two dates. Other phrasings almost certainly produce other shortlists.
- A single run per engine per date. We cannot separate twelve-week change from ordinary run-to-run variance.
- No independent verification of review counts or ratings. We recorded what the engines reported.
- No control for location, personalization, or account history on any engine.
- Some engines included businesses outside Andover proper without saying so, inflating disagreement by an unknown degree.
What survives those limits: one ordinary question produced eleven businesses across four assistants, with half the matrix empty. The reasons for the divergence are legible, and they can be measured.
The ASK Index
AI is already answering questions about your brand.
Find out which engines surface you, what sources shape the answer, and where the gaps begin.