Skip to content

The AI Recommendation Gap 2026 Method

How the US cells were measured

The fifteen US cells of the study: five trades in three cities.

Fieldwork October 20, 2026, in one night. Alfred sells services in this area. Alfred’s own products and prices are not part of any comparison in the study.

Cities and radii

A cell is one trade in one city: 15 cells, five trades in Houston, Chicago and Phoenix. The listings were counted inside a circle around each city center; the radius follows the size of the city, so the number of listings differs between cities while the shares need not. The Google results were asked for at city level, with the location string below.

CityCenterRadiusLocation for the Google results
Houston29.7604, −95.369825 kmHouston, Texas, United States
Chicago41.8781, −87.629820 kmChicago, Illinois, United States
Phoenix33.4484, −112.074025 kmPhoenix, Arizona, United States

In all, the US cells cover 10,348 category-listings, 33 AI answers and 203 first-run picks, collected on October 20, 2026. The design, the matching rule and the source classes are the same in every market and are set out in full on the cross-market method page.

Trades, categories and searches

Each trade is defined by one Google category. A listing counts when the category is its primary category or one of its additional ones, so a firm with two trades counts once in each. The Google search is the query below followed by the city’s name; the AI question uses the noun below.

TradeGoogle categoryGoogle searchNoun in the AI question
Plumbersplumberplumber + cityplumbers
Electricianselectricianelectrician + cityelectricians
House painterspainterhouse painter + cityhouse painting companies
Roofersroofing_contractorroofer + cityroofing companies
HVAC contractorshvac_contractorhvac company + cityHVAC companies

HVAC takes the place the builder trade has in the UK and Sweden, with hvac_contractor as its category. Figures across markets are compared on the three trades every city shares: plumbers, electricians and painters.

Listing counts

Listing data came from a business listings search, one count per measure and cell. The measures were all listings with the category in the circle, unclaimed listings, listings that link to a website, and listings with at least 1, 10 and 100 Google reviews.

The counts are category-listings, not distinct firms: a firm listed in two trades counts once in each, and closed listings are included. The Basics Score is the mean of three shares, with every trade pooled per city: listings that are claimed, that link to a website and that have at least ten Google reviews.

Google results and the top three

For each cell, Google’s organic search was asked for the trade’s query and the city, at city level, on desktop, ten results deep. Each search recorded whether an AI Overview appeared, the businesses in the map results with their rating and review count, the “People also ask” questions and the domains of the organic top ten. The “top three” means places 1–3 of the map results. AI Overviews load separately, so whether one appeared was recorded and its text was not.

The AI answers

The question, in English: “Who are the best {plumbers | electricians | house painting companies | roofing companies | HVAC companies} in {city}?”

  • ChatGPT: collected from ChatGPT’s consumer interface, logged out, location set to United States, language English. The answer carries named businesses with the rating and review count it shows, and its sources. The model is not always reported.
  • Perplexity: the sonar-pro model through its API, with web search on. The API takes no location.
  • Gemini: returned no answers and is not part of any figure.

Each cell had one ChatGPT answer and one Perplexity answer. The plumber question was put to ChatGPT a second time in each city, the re-run. Every named business was recorded, up to ten per answer, with the rating and review count shown and every cited link. Approximate figures such as “500+” were recorded as their lower bound.

Matching rule

To compare an AI pick with Google’s top three, both names are normalized: lower case, accents and punctuation removed, legal forms such as Ltd, LLC and Inc dropped. The city’s name is removed only when it is a whole word at the start or end and a distinctive word remains. Two names are the same business when their token-set similarity is at least 85 and they share at least one distinctive word, never only words such as “Plumbing”, “Electrical” or a place name. A name with no distinctive word matches only an identical name.

Google figures for picks

A pick’s Google rating and review count come from the top three, from a lookup by name (up to three per cell, chosen selectively) or from Google panels noted during fieldwork. “Best available” means the Google figure where it was verified and the figure the answer showed otherwise. A figure shown in an answer “matches” Google when it is within 10% or five reviews. One Phoenix lookup was excluded because it returned a different listing from the one on ChatGPT’s card; lookups that returned a branch of a multi-location firm are flagged.

How cited sources are classed

Every cited domain was classed as a directory, a review platform, a marketplace, a “best of” list, the business’s own website, Google Maps, social media or other. A cited site “does not rank on Google” when its domain was not in the organic top ten for the same search in the same city. That measure covers the aggregator domains (directories, review platforms, marketplaces and lists) and leaves out business sites and Google itself.

Limits

  • One question per assistant and cell, plus one re-run per city. Answers change between runs, so the results describe patterns, not any single business’s place.
  • ChatGPT was read logged out through a collection service and Perplexity through its API without a location; real users may see something else.
  • Lookups were selective and by name. Verified Google figures exist for 99 US picks.
  • Listing counts double-count firms in several categories, include closed listings and use radii that differ by city.
  • AI Overview text was not captured. Gemini is not included.
  • One snapshot, taken on October 20, 2026. Results and profiles change all the time.

Data about individual businesses is not released, and a business named in the study appears only in a neutral or positive context.

The cells

CityTradeCategoryRadiusListingsChatGPT picksPerplexity picksRe-run picks
HoustonElectricianselectrician25 km79557–
HoustonHVAC contractorshvac_contractor25 km94458–
HoustonHouse painterspainter25 km51458–
HoustonPlumbersplumber25 km711595
HoustonRoofersroofing_contractor25 km1,15368–
ChicagoElectricianselectrician20 km61158–
ChicagoHVAC contractorshvac_contractor20 km60169–
ChicagoHouse painterspainter20 km340510–
ChicagoPlumbersplumber20 km505575
ChicagoRoofersroofing_contractor20 km64758–
PhoenixElectricianselectrician25 km71257–
PhoenixHVAC contractorshvac_contractor25 km91168–
PhoenixHouse painterspainter25 km45968–
PhoenixPlumbersplumber25 km830595
PhoenixRoofersroofing_contractor25 km615510–

Source: Alfred, The AI Recommendation Gap 2026, United States, cell table read from the study’s cells file, October 20, 2026.

How to cite this method

Alfred, The AI Recommendation Gap 2026, United States: method, October 2026, https://tryalfred.ai/us/reports/ai-recommendation-gap/method

Quote the figures with credit to Alfred and a link to this page. Data about individual businesses is not released.

The same measures, for your business

Check my business reads your own listing and the map from 49 points, and measures each line against a fixed target.