Four AI engines. The same 372 businesses. Most make no one’s list, and the engines rarely agree on who does. Here’s what every AI rewards, where they split and how fast it all moves. The third reading of a longitudinal study of the same 372 businesses across legal, medical, restoration, home services and franchise verticals, and the first reading across four AI engines: Perplexity, ChatGPT, Gemini and Claude. Readings: March, July and October 2026.
Volume 3 at a Glance
Volume 1 asked whether AI could find these businesses. Volume 2 showed it finds almost all of them and recommends very few. Volume 3 put the same 372 businesses to four AI engines. The gap showed up on every one, and each engine picked different winners.
The Finding That Matters Most
There is no single “AI ranking” to win. There are four shortlists, and they barely overlap. But the businesses shut out of all four share the same traits, and so do the ones that make several. That’s the part you can act on.
Seven Things We Found:
1. The gap is universal. Every engine could describe 77% to 81% of these businesses and recommended 9% to 16%.
2. The same businesses get left out everywhere. 287 of 372 made no engine’s list. The bottom third on our company score was recommended 4.6% of the time across ChatGPT, Gemini and Claude; the rest, about 20%.
3. Engines pick different winners. Only 21 businesses were recommended by all four.
4. Perplexity is the stingiest, and it just changed. About 2 names per buying answer. In late September it switched systems, and its answers shrank 52% to 74%.
5. Sources differ, with two constants. Yelp drives Claude, LinkedIn drives Perplexity, ChatGPT leans on businesses’ own sites. BBB and Super Lawyers are in every engine’s top 10.
6. The name problem is everywhere. Even with the company named, engines found the right owner 55% to 64% of the time. Franchise owners: about a quarter.
7. AI forgets fast. On Perplexity, half of what AI cites about a business is replaced in about six weeks.
Why You Should Care
When a buyer asks AI who to hire, the answer is the shortlist. In October, 77 of every 100 businesses in our panel were on no engine’s shortlist, including most of the ones every engine could describe. If you aren’t named, you aren’t considered and most analytics won’t show you that it happened. Invisible owners create invisible companies. Now there are four rooms you can be invisible in.
How to Read This Report
Volume 3 combines two kinds of evidence. One goes deep on a single engine over time. The other goes wide across four engines at once.
Five Terms Used Throughout
• Known: AI names the company when asked what it is.
• Recommended: AI names the company when a buyer asks for the best, or who to call for a problem.
• Right person: asked who the owner is, AI describes the right person, tied to the right business and city.
• Company score: the index’s 0 to 100 measure of how many earned outside sources AI cites about a company and how authoritative they are.
• n: the number of businesses a figure is based on. “Legal (n=80), recommended 11%” means 11% of the 80 law firms in the study, about 9 firms. The bigger the n, the more weight the number can carry: a difference of a few points means more across 372 businesses than across 56. We show n on every vertical figure so you can judge that for yourself.
| Why Perplexity Shows Two Recommendation Figures Our own October reading on Perplexity (12 October, six questions) found 13% of companies recommended. Thrive’s four-engine reading three days later, on Perplexity’s pinned fast preset, found 9.1%. Both are real and both are labelled. Cross-engine pages use the 15 October figure so all four engines share one method. |
| Different Rulers, On Purpose Volume 2 reported 86% of companies known and about 19% recommended. This report applies one stricter definition to both July and October, so its figures run lower: 78% known and 16% recommended in July. Compare figures within this report, not across volumes. Volume 1’s 46% and Volume 2’s 9% measured different things and are not a trend. |
Two Question Sets on Perplexity
Every business was measured twice. Arm A asks the exact questions from March, so it carries the year- long trend. Arm B asks cleaner, more natural questions, introduced in July. Trends use Arm A. Where Arm B tells a different story, we show both.
Known Everywhere, Picked Nowhere
Same 372 businesses, same five questions, four engines. Every engine could describe most of them. None recommended more than one in six. 198 were known by all four and recommended by none.
The Gap Is the Rule, Not the Exception
Being known is table stakes on every engine. Being picked is rare on every engine. Restoration felt it most: 42 of 56 firms were known by all four and recommended by none.
WHY IT MATTERS
What Gets You Picked
The engines pick different winners. But the businesses they all pass over look alike. We tested one question: does a weak outside reputation keep you off every engine’s list?
| How We Tested It 1. Every business already has a company score (0 to 100) from our Perplexity reading. It measures how many credible outside sites AI cites when describing the company, and how authoritative they are. 2. We split the 372 businesses into three equal groups of 124: weakest scores, middle, strongest. 3. We counted how often each of the four engines recommended the businesses in each group. |
How To Read it
In the weakest group, every engine recommended 2 to 7 businesses out of 100. In the middle and strongest groups, 12 to 23 out of 100. Only 11 of the 124 weakest businesses made any engine’s list. The score was built from Perplexity’s answers, yet it sorted businesses the same way on ChatGPT, Gemini and Claude: every engine leans on credible outside sites that describe you.
Both comparisons show an association, not proof of cause. Directory presence is counted from the directories the engines cited when asked what the company is. Ahrefs Domain Rating showed little to no relationship with being recommended on any engine.
WHY IT MATTERS
The 21 on Every List
21 businesses were recommended by all four engines. They aren’t bigger, and they don’t have stronger websites by the usual SEO measures. Four things set them apart from the 287 that no engine recommended.
1. Less crowded markets. For the 21, the engines named a median 25 other businesses across these questions; for the shut-out group, 46. Split the panel by crowding and the pattern is stark: in the least-crowded third, 39% made at least one list; in the most-crowded third, 10%. 17 of the 21 sit in that least-crowded third, mostly in smaller cities and suburbs.
2. Medical, then everyone else. 10 of the 21 are medical practices. The other 11: franchise 4, home services 3, legal 2 and restoration 2. As a share of each vertical, medical 12.5% (10 of 80) made all four lists, franchise 5.0% (4 of 80), home services 3.9% (3 of 76), restoration 3.6% (2 of 56) and legal 2.5% (2 of 80).
3. Above the line. All 21 sit in the middle or top third on company score. None came from the bottom third.
4. Their own site does more of the talking. When the engines described them, 38% of citations pointed at their own website, against 30% for the shut-out group. A site that says clearly what you do and where gives every engine the same answer.
What Didn’t Separate Them
Size, domain rating, the number of directories and schema markup looked the same in both groups.
You can’t choose how crowded your market is. Everything else on this page you can work on.
WHY IT MATTERS
The Name Problem Is Universal
On every engine, AI struggles to tie owners to their businesses. Even with the company named in the question, the four engines found the right person, business and city 55% to 64% of the time, and franchise owners about a quarter of the time. On Perplexity we went further. Give it an owner’s name and company, and it almost never gets the wrong person: 3 of 372. Give it only the name, and it found the right person, business and city 30% of the time in October (111 of 372), down from 40% in July.
We read the wrong-person answers by hand. They describe a fitness podcaster, a university professor, an Australian fashion curator: real people who share a name with the owner. Wrong-person answers almost never cite the owner’s own site (0.1 citations per answer, against 1.2 for correct ones).
The Formal-Name Fingerprint
When AI writes “Mark D. Stubbs,” it is usually reading an authoritative listing, and it is usually right: 89% of formal-name answers were accurate, against 63% of the rest. A middle initial doesn’t make you visible. One consistent name, tied to your business and city everywhere AI looks, does.
WHY IT MATTERS
Four Engines, Four Different Lists
The engines agree that most businesses don’t make the cut. They don’t agree on which ones do. Ask two engines the same buying question and only 6% to 11% of the businesses they name overlap, depending on how strictly names are matched. 80% to 87% of the businesses named appear on just one engine.
How Long Is the Shortlist?
Perplexity names about two businesses per buying answer. Claude names about eight.
Median number of businesses named in answers to the three buying questions, by engine (about 1,116 answers per engine)
Same patterns, every engine. Restoration is recommended least everywhere (4% to 7%). Legal firms are the least known everywhere (58% to 60%). Franchise owners are the hardest to identify on every engine: about a quarter found as the right person, even with the company named.
ChatGPT, Gemini and Claude agree with each other more than with Perplexity. Of the businesses either engine recommended, ChatGPT, Gemini and Claude shared 50% to 56% pairwise; with Perplexity, 35% to 47%. The stingiest engine is also the odd one out.
WHY IT MATTERS
Every Engine Reads Different Sources
Ask why the engines disagree and the answer is in their citations. Each one leans on a different set of websites to decide who’s worth naming.
Four personalities. Perplexity reads directories and LinkedIn. ChatGPT reads fewer sources (four per answer) and leans hardest on the business’s own site. Gemini spreads out, and still cites Facebook, Reddit and YouTube. Claude reads the most, with Yelp in 61% of its answers, and cites data brokers like ZoomInfo more than any other engine.
The Two Universal Sources
Only two sites made every engine’s top 10: BBB, which spans every category, and Super Lawyers, which carries legal. If you do one listing this quarter, that’s the evidence for where to start.
WHY IT MATTERS
The Year So Far on Perplexity
Read the three volumes together and the year looks like an accordion. Between March and July, AI opened up: longer answers, more sources, higher scores for nearly everyone. Between July and October, it closed again.
• Volume 1 (March) found AI knew companies better than the owners behind them, and nearly half of owners
were effectively invisible. The question was visibility.
• Volume 2 (July) found visibility was never the point. AI could find almost everyone. What was scarce was being
recommended.
• Volume 3 (October) finds the scarce thing got scarcer. AI tightened its answers, and fewer businesses made the list.
What held all year: the gap between being known and being recommended. Perplexity knew most of these businesses in every reading and recommended a small minority. In October, three more engines confirmed the same gap (Part 2).
ACROSS THE OTHER ENGINES
This year-long view exists only for Perplexity. In October, ChatGPT, Gemini and Claude showed the same end state: they knew 78% to 81% of these businesses and recommended 14.5% to 15.9%. Their trend lines start with the next reading.
The List Got Shorter
We asked the same six questions about the same 372 businesses, word for word. In July most answers ran into the 800-token limit. In October none came close.
• A 10-source ceiling appeared. In October, 2,228 of 2,232 answers cited exactly 10 sources. In July they averaged 15.7, with no fixed limit. Trimming July to its first 10 sources moves only 1 of 372 businesses to a new quadrant, so what hurts is which 10 AI now picks.
• AI made up its mind. Asked twice, six hours apart, owner answers matched for 41 of 50 businesses. In July it was 25 of 50. Part of that is built in: one search and 10 fixed slots leave less room for answers to vary.
Disclosure: Perplexity Replaced the System Behind Our Questions
On 27 September 2026, between our July and October readings, Perplexity retired the API we use (Sonar Chat Completions) and began re-routing those requests to its newer Agent API. Its named replacement for sonar-pro, the “fast” preset, runs one search capped at 10 results. Our prompts went through word for word; the system answering them was different.
ACROSS THE OTHER ENGINES
The ten-source cap is Perplexity’s alone: 89% of its answers cited exactly 10 sources, against 3% for ChatGPT and Gemini and 28% for Claude. Its answers were also the shortest: a median 93 words, against 191 (ChatGPT), 256 (Gemini) and 337 (Claude).
WHY IT MATTERS
Who AI Reads Now
AI didn’t just cite less. It changed who it listens to. Community and social sources nearly disappeared. Structured directories and marketplaces filled the space. Arm B shows the same shift (directories 23% to 33%).
Who Took It
New in October’s top 15 cited sources: Zocdoc, WebMD, Thumbtack, Houzz, HomeAdvisor, Best Lawyers, FindLaw, Best Law Firms and MediFind.
ACROSS THE OTHER ENGINES
Directories are the biggest earned source on every engine (51% to 83% of earned citations). But the exit from social is Perplexity’s: Gemini still cites Facebook, Reddit and YouTube, and Claude cites Facebook in 11% of answers..
WHY IT MATTERS
The Six-Week Memory
Ask AI the same question twice in one morning and 87% of the sources repeat. Ask in July and again in October, and only 25% do (24% on Arm B). That holds even when we compare just the first 10 sources from each wave.
Half-life is how long until half the cited sources are replaced, after allowing for normal same-day churn. Turnover sped up about threefold after July.
First 10 citations per answer in each wave. Wave 1 date approximated as mid-March 2026.
Visibility Is a Subscription, Not a Purchase
A press mention gets you into the answer. Within weeks, AI swaps it out. Your own site and your directory profiles are what stay. The businesses that stay on the list keep feeding it.
ACROSS THE OTHER ENGINES
Turnover hasn’t been measured on ChatGPT, Gemini or Claude yet; the next reading adds a same-day repeat on every engine. One clue: answering the same question about the same business, two engines share only about 1 in 7 of the sites they cite. What AI reads about you depends on which AI you ask.
WHY IT MATTERS
Fewer Names on the List
The share of businesses AI names when a buyer asks for the best, the most trusted or help with a problem fell on both instruments.
This is not noise. Asked twice in one day, AI gave the same recommend-or-not answer for 48 of 50 companies. Between July and October on Arm B, 43 companies lost their place on the list and 14 gained one.
The Receipt, Not the Reason
When AI does recommend a business, it links to that business’s own website in 55% to 83% of those answers. The site is how AI proves its pick. It is rarely why AI picked it.
ACROSS THE OTHER ENGINES
In October, ChatGPT, Gemini and Claude recommended 14.5% to 15.9% of these companies, against 9.1% for Perplexity on the same five questions. Restoration trailed on all four engines (page 11).
WHY IT MATTERS
Companies Slipped, Owners Held
Company visibility fell about 2.5 to 3 points on both instruments, and losers outnumbered gainers about 4.5 to 1. Owner visibility held flat.
How We Count a Name
Our scoring rule looks for an owner’s exact name in the AI’s answer. In October, Perplexity started writing many professionals’ names the way formal directories list them, with a middle initial: “Mark D.
Stubbs” instead of “Mark Stubbs.” The answers were correct, but an exact-match rule counts them as misses.
We keep the original rule as the official series and report a re-score that accepts middle initials (+0.3 instead of -5.4), because it reflects what AI actually said.
ACROSS THE OTHER ENGINES
Every engine favors the company over the person. Asked to recommend “someone with a strong reputation,” the four engines named the company in 5% to 12% of answers and the owner in 1% to 4%.
WHY IT MATTERS
Business or Machine?
When a score moves, either the business changed or the machine did. We pulled Ahrefs link and traffic data for every panel domain to test the business side. It came back nearly empty.
Quality link: linking site DR 30 or higher, still live, followed, excluding link-farm domains (any site linking to 10 or more of the 34 unrelated mover businesses). Link counts are medians for the 34 mover domains. Traffic is Ahrefs’ estimate including www, for sites with July traffic.
Domain Rating didn’t move, and didn’t predict movement. Median DR change was zero in every group. Across all 372 businesses, DR change and company score change are unrelated (r = 0.00). Persistent gainers came in with somewhat stronger sites, but DR level alone doesn’t predict who moved either (r = -0.02).
Quality links leaned the right way, weakly. Persistent gainers added a median 13.5 quality links after July, persistent losers 7. Business by business, link count barely tracks score change (r = 0.11, n=34). Every group added more links after July than before, losers included.
Traffic fell for everyone. Estimated organic traffic dropped 22% at the median from July to September (n=335), with 79% of sites down and every vertical between -19% and -25%. Gainers and losers fell alike. That’s the measurement environment, not AI.
ACROSS THE OTHER ENGINES
The four-engine reading points the same way. The same 372 businesses, with the same websites and the same links, got different results on each engine: only 21 made all four lists, and Domain Rating showed little to no relationship with being recommended anywhere. When the business holds still and the answer changes, the machine is doing the moving.
THE VERDICT: MOSTLY MACHINE. WHY IT MATTERS
Mover Stories
Five anonymized case notes: three of the 20 companies that rose in both intervals on Perplexity, two of the seven that fell in both, and how the other engines treated them in October.
ROSE IN BOTH INTERVALS
Company score: March 72.5 · July 81.2 · October 96.2
What changed. 62 quality new links after July, up from 13 between March and July, one of the three largest jumps among the 34 movers. Press coverage appeared in its citations for the first time.
Other engines, October. Known by all four engines in October, recommended by none. Strong sources lifted its Perplexity score; they didn’t put it on a shortlist.
What it means. Business work and score lined up on Perplexity’s measure. Getting named on a shortlist takes more than strong sources.
ROSE IN BOTH INTERVALS
Company score: March 7.0 · July 53.3 · October 65.0
What changed. No real link growth: 6 quality new links after July, 5 before. What changed was where AI looked. Earned sources went from none of its citations to 20%, including its first press mention.
Other engines, October. Known by all four engines, recommended by none.
What it means. A small B2B firm that went from nearly invisible to known in a year. One outside mention can claim one of 10 slots.
ROSE IN BOTH INTERVALS
Company score: March 52.9 · July 67.5 · October 71.2
What changed. Almost nothing on the web: domain rating 0 and two quality new links after July.
Other engines, October. Recommended by ChatGPT only.
What it means. It rose without touching its footprint. When the whole answer pattern shifts, some businesses rise by default. Treat a gain like this as luck, not proof.
FELL IN BOTH INTERVALS
Company score: March 75.0 · July 68.8 · October 65.0
What changed. Strong domain rating (36 to 37) and steady new links. In July, Perplexity recommended it, citing six platforms. In October: not recommended, two platforms, and its own site carried half the citations.
Other engines, October. Recommended by Gemini and Claude, not by ChatGPT or Perplexity. Slipping on one engine didn’t mean slipping everywhere.
What it means. Its links didn’t fail it. The shorter list did. With 10 source slots, a business whose proof lives on its own pages loses the outside voices that vouch for it.
FELL IN BOTH INTERVALS
Company score: March 71.9 · July 68.8 · October 65.0
What changed. 19 quality new links after July, the most of the seven persistent losers, up from 4. Still recommended in October, but its supporting sources shrank from six platforms to two, and earned sources from three to one.
Other engines, October. One of only 21 businesses in the panel recommended by all four engines.
What it means. Link building kept it on the shortlist. It didn’t stop the slide. Breadth of outside mentions is what the machine is cutting.
The Five Verticals
Each card shows four numbers. Company score is the original-rule mean in March and October (Arm A). AI knows them is the share AI can describe when asked directly. Recommended is the share AI names in buying answers (Arm A; Arm B in the text where it differs). Owner found by name counts owners AI matched to the right business and city from their name alone. The four cells are the Perplexity series; the line below each card adds the four-engine reading.
Legal is still the hardest place for a firm to be known: AI recognizes just 57% of these firms. Lawyers, though, are the easiest owners to identify by name, because bar directories list them formally.
October’s legal answers lean on Super Lawyers, Best Lawyers, Avvo, FindLaw and Lawyers.com. Justia, a July favorite, fell 86%.
Across four engines: known 58% to 60%, recommended 5% to 16%, owner found as the right person 65% to 75%. 29 of 80 known by every engine and recommended by none.
What it means: the legal directories are the shortlist. Being listed, complete and consistent on them is the work.
Medical keeps the highest recommendation rate in the panel, but it fell from 35% to 25% on the cleaner instrument. Zocdoc and WebMD now carry the medical answer, followed by MediFind and Healthline. Healthgrades, a July staple, dropped 79%.
Across four engines: known 79% to 86%, recommended 14% to 24%, owner found as the right person 72% to 80%. 39 of 80 known by every engine and recommended by none.
What it means: the booking and health directories are where AI decides. A practice missing from Zocdoc is missing from the answer.
Known almost everywhere, recommended almost nowhere. Restoration answers lean on BBB, Thumbtack, Houzz and HomeAdvisor, with national brand sites among the most cited of the rest.
Across four engines: known 84% to 86%, recommended 4% to 7%, owner found as the right person 42% to 77%. 42 of 56 known by every engine and recommended by none.
What it means: in a category built on emergencies, AI hands the call to the names on the marketplace lists.
The biggest company decline in the panel (-5.7, and -7.1 on Arm B), and recommendations nearly halved on the cleaner instrument (25% to 13%). BBB, Thumbtack, ServiceAgent and Houzz dominate.
Across four engines: known 83% to 91%, recommended 13% to 24%, owner found as the right person 53% to 74%. 35 of 76 known by every engine and recommended by none.
What it means: the marketplaces now answer the trades question. Show up there, with reviews or AI answers without you.
The brand without a face. AI knows the brand, and almost never the local owner: 5 of 80 found by name, 47 confused with someone else. On the cleaner instrument, recommendations fell from 38% to 29%.
Across four engines: known 83% to 85%, recommended 9% to 13%, owner found as the right person about a quarter. 53 of 80 known by every engine and recommended by none.
What it means: the local owner is the open seat. A named, findable franchisee is the cheapest differentiator in the system.
What This Means If You Run One of These Businesses
You can’t win “AI” as one thing. There are at least four shortlists, and they barely overlap. But the price of admission is the same on all four, and most businesses haven’t paid it.
1. Get out of the bottom third. Businesses with the weakest outside footprint were recommended 4.6% of the time across ChatGPT, Gemini and Claude. Credible outside sources that describe you are the entry ticket on every engine.
2. Get on one directory AI reads. Start with BBB. Businesses with no directory cited were recommended about half as often on ChatGPT, Gemini and Claude. BBB is the one cross-category source all four cite heavily. Then add the list that rules your category: Super Lawyers, Zocdoc, Angi or Houzz.
3. Know your customers’ engine. Yelp is in 61% of Claude’s answers and none of ChatGPT’s. LinkedIn drives Perplexity. ChatGPT leans on your own site. Check the one your buyers use.
4. Treat your website as the receipt. Every engine cites the business’s own site in most answers that recommend it (60% to 74%). It won’t get you picked. It proves the pick.
5. One name, everywhere. Even with your company named, engines found the right owner only 55% to 64% of the time. Use one name form, next to your business and your city, everywhere AI looks.
6. Refresh every six weeks. On Perplexity, half of what AI cites about you is replaced in about six weeks. A listing updated once a year is a listing AI forgets.
| What Won’t Move the Needle More links, on their own. Domain rating showed little to no relationship with being recommended on any engine. And past the first directory or two, piling on more listings didn’t raise the odds. Presence beats volume. |
In Plain English
Every AI starts from the same pool: businesses it can verify through credible outside sources, at least one directory and a working website. From that pool, each engine picks its own favorites. Get into the pool first. Then show up where the engine your buyers use is looking.
The One thing to Take Away
Invisible owners create invisible companies. Four engines checked. 77 of every 100 businesses were on no one’s list. Make sure your name is on at least one, spelled the same way everywhere AI looks.
What This Means for Marketing Teams
If you run marketing for a local business, or a portfolio of them, October broke seven assumptions that are probably still in your plan.
Your First 30 Days
1. Ask the buyer’s question on two engines. Run “best [category] in [city]” and “who should I call for [common problem]” in the two AI tools your buyers use most. Note who gets named and which sites get cited.
2. Find your client’s directories. Start with BBB, then the sites cited most in those answers. Audit each profile for completeness, accuracy and recent reviews.
3. Fix the owner’s name. One name form, everywhere, next to the company and city. Ask each engine “who is [name] at [company]?” and log who comes back
4. Set the cadence. A six-week refresh on the calendar, and a quarterly re-test on every engine you track.
| What We Can’t Tell You Yet Four engines read once show what’s true now, not how each one moves. Three Perplexity readings show change, not cause. We don’t yet test follow-up questions, where buyers narrow the list further. The next reading starts every engine’s trend line. |
Where the Index Goes Next
Volume 3 turns the Business Visibility Index from a study of one engine into a study of AI. From Q4, every reading covers the same 372 businesses on every major engine, so we can separate what’s structural from what’s one engine’s quirk.
What Every Reading Will Measure, on Every Engine
• Known, recommended and right person, by vertical, always showing how many businesses each figure is based on.
• Shortlist length and agreement: how many businesses each engine names, and how often they name the same ones.
• Source maps: which websites each engine cites, and how fast they turn over.
• Instrument control: pinned models and settings, logged versions and a same-day repeat on every engine, so platform changes are visible instead of hidden.
What We’re Still Digging Into
• Google AI Overviews. The fifth engine, added next reading.
• A live test. In Q4 we’re placing tracked content for a small, separate group of businesses (owned, managed and earned) and timing how long each engine takes to notice.
• Anchors. A set of well-known national businesses, measured from now on, so future waves can separate platform drift from real change.
• Follow-up questions. How the list narrows when a buyer asks about price, services or specialties.
Why It Matters
A business owner doesn’t care which engine their customer uses. They care whether they’re on the list. Tracking what’s structural tells them what to fix once. Tracking what’s engine-specific tells them where to look next.
How We Measured This, in Plain Language
The Perplexity series. We asked Perplexity (sonar-pro, through its API) six questions about each of the same 372 businesses and their owners: who the owner is, who the well-known experts are, who to trust, what the company is, who the best companies are, and who to call for a problem. We did this in March, on 30 and 31 July, and on 12 and 13 October 2026, with identical wording each time, verified word for word before any question was sent.
The platform changed underneath us. Perplexity retired the API behind every BVI reading on 27 September 2026 and now re-routes those requests to its Agent API, where its recommended replacement for sonar-pro runs one fast search capped at 10 results. The October reading used the same endpoint, model name and prompts as July, so the system changed but our questions didn’t. Independent check: on 15 October, Thrive re-ran 1,860 of the same prompts (five per business) on Perplexity’s Agent API with the fast preset pinned. It reproduced our reading: companies described 77% vs 78%, owners identified 97% vs 97%, 72% of cited pages identical, and the same recommend-or-not result for 346 of 372 businesses. The pinned run recommended slightly fewer companies (9% vs 13% on the two buying questions).
Ahrefs data. Ahrefs’ October referring-domain totals were set aside: they ran about 2.5 times August’s counts for the same sites (median 1,130 vs 471, n=372). A link-farm surge explains it: new linking domains for the 34 movers jumped from about 4,500 in August to 25,457 in September, 76% from sites below DR 5, many on .shop and .store domains linking to all 34 businesses at once. Domain rating matched August closely (r = 0.98). Organic traffic after July 2026 is provisional: Google’s anti-scraping changes since September have degraded third-party ranking data, and a re-pull revised May to July panel totals down 4% to 6%.
Four engines. On 15 and 19 October, Thrive put five of the six questions (all but “well-known experts”) for all 372 businesses to Perplexity, ChatGPT, Gemini and Claude, and scored every answer with the rules used for the Perplexity series. See Parts 2 and 3.
Two instruments. Arm A repeats the March questions exactly and carries the trend. Arm B uses cleaner wording and more accurate business types, introduced in July. Both ran on the same days.
Noise floor. Fifty businesses were measured twice in one day, about six hours apart, to see how much AI varies on its own.
Scoring. Each score runs 0 to 100: half for how many earned outside sources AI cites, half for the authority of the best one. Owned sites, data brokers and people-finder sites earn nothing. The rule is unchanged since March.
Recommended means AI named the business or owner in its answer to a buying question. Known means AI described it when asked directly. Source types (owned, managed, earned) are assigned by domain. Half-life compares the exact pages cited for the same question in each wave.
Two owner numbers. October owner figures are reported with name-tolerant matching (page 19). The original exact-name series is shown beside it. March cannot be re-scored, because its raw answers were not retained.
Disclosures, in Brief
1. Trend figures come from one AI platform, Perplexity sonar-pro. Parts 2 and 3 use a single cross-engine reading (Perplexity Agent API fast preset on 15 October; ChatGPT gpt-5, Gemini 2.5 Flash and Claude Sonnet 5.5 via their APIs on 19 October), five questions per business, one pass, no same-day repeat. API answers can differ from the consumer apps. Gemini’s redirect links were mapped to the source domain it reported; 13 blank Gemini answers were excluded. Names per answer were counted by automated extraction. Google AI Overviews was not measured.
2. On 27 September 2026, Perplexity retired the Sonar Chat Completions API used for every BVI reading and began re-routing those requests to its Agent API. The October reading ran on the same endpoint, model name and prompts, but on a different system: shorter answers, a ten-source cap, a different source mix. Perplexity’s migration guide maps sonar-pro to its “fast” preset (one search step, Fast Search, max_results 10, an OpenAI model with reasoning off, per its presets documentation); the settings applied to re-routed requests are not visible to us, but a 15 October re-run with the fast preset pinned reproduced the October reading closely (page 29). Movement after July reflects the platform as well as the businesses. From Q4, readings will call the Agent API directly with the preset and max_results pinned, so the instrument is ours.
3. No national anchor businesses were measured in any wave, so platform-wide drift cannot be subtracted from real change.
4. In October, Perplexity began writing many owners’ names with middle initials. The exact-name rule undercounts them; we report a name-tolerant re-score beside the official series. About 5% of July identity answers had the same issue.
5. Arm B has no same-day repeat of its own; its changes are judged against Arm A’s noise floor.
6. Source types are assigned by domain rules. “Other sites” (28% to 44% of citations) is an unclassified long tail.
7. Recommended, known and accurate are detected from names in the answer text. Accuracy categories were spot-checked by hand, not fully hand-coded.
8. Half-life assumes the March reading took place in mid-March 2026.
9. Volume 1 figures are quoted from the corrected verification ledger, not the original Q1 report. Volume 2 and Volume 3 figures use different definitions and are not compared directly.
10. Organic traffic estimates after July 2026 are provisional because of Google’s anti-scraping changes. Ahrefs’ October referring-domain totals were not used (inflated by a September link-farm surge); link analysis counts only DR 30+, live, followed links and excludes link-farm domains.
11. The 372 businesses are the March panel of 400 minus 28 records outside their category. No business was contacted or influenced.
12. Three readings begin to separate trend from noise. They do not establish cause.
13. Company-score thirds use Perplexity Arm A October scores, cut at 61.2 and 65.0. Directory presence counts distinct directory sites cited across the four engines’ “what is [company]” answers.
Business Visibility Index, Volume 3, Q3 2026 · Thrive Internet Marketing Agency · © 2026. All rights reserved.
