Which Marketing Contractors ChatGPT and Claude Recommend: What Moved in a Week

Who moved, by name

ChatGPT's ten lost Chief Outsiders and gained KlientBoost and Bop Design. Claude's ten lost Powered by Search and Walker Sands and gained Siege Media, First Page Sage and Belkins, which rose from 8 recommendations to 14, the largest rise of the three. Every company that entered a ten had already been recommended in week 1. Because of tied scores at tenth place, both tens now hold eleven companies.

Claude also recommended less: about 12% fewer recommendations than a week earlier, across 182 companies instead of 206. ChatGPT's total stayed level, even with two fewer answers.

A week against a repeat

Issue 1 said we would not call a rise or a fall a trend until a week moves the list more than simply asking the question again does. Here is that comparison, counted the same way on both sides.

For ChatGPT the week moved the list less than a repeat does. A single appearance or disappearance is not enough to say that a company's position changed.

For Claude the week moved it about five points more. We will call that a change only if it holds in the coming weeks.

Both assistants now link to Clutch first

In week 2 the directory clutch.co is the most linked site in the answers of both assistants: 44 links in ChatGPT's answers and 26 in Claude's, where it had 14 a week earlier. A week earlier Claude linked most often to Directive's own website. Behind clutch.co, both link mainly to the agencies' own sites.

A link in an answer is what the reader can click. It is not a record that the model read that page.

Site access barely moved

We also check whether the recommended companies' websites let AI crawlers in. In week 1, thirteen of the 149 sites we could test refused at least one named AI crawler. In week 2 it was the same thirteen and one more. In those same seven days, 82 companies left the recommendations. The list changed a lot; access to the sites hardly changed at all.

These are the answers the sites gave to our own requests, which carry a crawler's name. It is not a record of what any assistant read.

What to do with this week

If you are choosing an agency. One answer shows only part of the list. Ask your question two or three times before you write any name down, and keep the names that come back each time. A name you saw once you may not see again.

If you are watching your own company. Being named once is not a position. Of the 82 companies that dropped out, 62 had been recommended only once in the whole week, and no newcomer this week got past two recommendations. What held was the top group: nine of ten names for ChatGPT, eight of ten for Claude. So measure how often you appear across repeats and across weeks, and count a rise as progress only when it holds week after week.

Where the buyer lands. If your name comes up, the page the buyer can click from these answers is most often a Clutch profile or your own website. Keep both up to date.

What a site check alone will not show. The same fourteen sites refuse crawlers, while a third to a half of the list changes every time anyone asks. A check of your robots.txt alone will not tell you why your name moved.

What we are watching next week

Every Monday we ask both assistants the same twenty questions, three times each. The next issue reports which companies entered and disappeared by name, how many of each assistant's own ten held, and whether a week again moves a list more than a repeat does.

Next measurement: 2 November 2026.

Where these numbers stop

Two weeks is two measurements. Everything above compares two Mondays.

These twenty questions. They are printed in issue 1, Appendix D. A question that describes a buyer and a question that names a category return largely different companies.

No cause. We measure who is recommended and how much it moves. Why a particular company is recommended is not established here.

No crawler behaviour. We send requests that carry a crawler's name from our own networks. That says nothing about the robots that build training sets, and nothing an assistant shows us records that it opened a page.

Measure this for your own market

Measure which providers ChatGPT and Claude recommend when buyers ask about your market, with repeats, so you can tell real movement from noise. See where your company appears, which competitors come back, and which websites the answers link to.

Get your market measured

Appendix: the counts behind the main text

ChatGPT recommended 90 companies in each week, and 64 of its week-1 companies were recommended again: 71%. Claude recommended 206 companies in week 1 and 182 in week 2, with 128 of the earlier set recommended again: 62%. Counted across both assistants, 157 of 239 companies were recommended in both weeks.

The leaders that moved are these. Chief Outsiders left ChatGPT's own ten (10 recommendations in week 1, 8 in week 2), and two companies entered it, both at 9: KlientBoost (8 in week 1) and Bop Design (8). Claude's ten lost Powered by Search (16 to 13) and Walker Sands (18 to 13) and gained three: Siege Media (12 to 16), First Page Sage (11 to 15) and Belkins, the largest rise among the newcomers, from 8 to 14. No company entered either ten from nowhere: every newcomer to a ten was already recommended in week 1.

What left and what arrived carried little weight. The 82 companies that disappeared accounted for 111 of the 1,087 recommending observations in week 1, about one in ten, and the largest single loss is Foundation Marketing at 6 observations, followed by Wpromote at 4. The 65 that appeared carry 79 observations between them, and none of them is above 2.

One assistant also became measurably less generous, and it is not the one that lost answers. ChatGPT made 347 recommendations in week 2 against 349 in week 1 — flat, despite answering two fewer questions. Claude made 648 against 738, down 12%, spread over 182 companies instead of 206. Across both assistants the week carried 995 recommendations against 1,087.

The two assistants also drifted slightly further apart: 57 companies were recommended by both in week 1, and 50 in week 2.

In week 1 the most linked domain in ChatGPT's answers was the directory clutch.co with 55 links, while Claude linked first to directiveconsulting.com. In week 2 clutch.co is first for both: 44 links in ChatGPT's answers and 26 in Claude's, where it had 14 a week earlier. Behind it, ChatGPT links to the providers' own sites — ironpaper.com 27, walkersands.com 26, kalungi.com 25, directiveconsulting.com 21 — and Claude to directiveconsulting.com 24, thegrowthsyndicate.com 23, beomniscient.com 19.

Four domains left Claude's ten most linked (poweredbysearch.com, theb2bplaybook.com, tripledart.com, firstpagesage.com) and four entered it (belkins.io, thecmo.com, growthspreeofficial.com, nogood.io). ChatGPT's ten changed by one domain: gorilla76.com out, beomniscient.com in.

We count the links shown in the answer separately from the addresses the assistants' own search tools returned.

The set of sites that refuse a named AI crawler is stable, unlike the recommendations. Thirteen of the 149 recommended companies whose website we could identify refused at least one named crawler on every attempt in week 1; in week 2 the same thirteen refuse again and one site joined them, so fourteen refuse — while 82 companies left the recommended list in those same seven days.

One site accounts for the change: stackmatix.com closed to named AI crawlers both in its robots.txt and in practice. And a stable headline number can still hide movement: 73 of the 149 sites published an llms.txt in week 1 and 73 in week 2, but not the same 73 — two sites added the file and two no longer serve it.

The response codes, the sites that limited how fast we could ask, and how much of the week the second network covered are in the appendix.

Appendix: how this issue measures what moved

The series prints the rules it measures by, and marks comparisons that cross a change in them. Six notes apply to this issue.

How a company is recognised between two measurements. Matching goes by website when both weeks resolved one, and otherwise by an exact company name that occurs exactly once on the other side; a name that matches more than one company produces no match and is published as unmatched. Of the companies recommended in both weeks, 105 matched by website and 52 by name, the latter being companies whose website resolved in only one of the two weeks. Website-only matching would count such a company as both gone and new, which is why the name step exists.

One unit for both kinds of movement. Spread inside a week and movement between weeks are both measured as one repeat against one repeat, over every ordered pair of repeats, where an observation is one company put forward for one question. Measured this way, week 1 gives the same figures this series published for it: 428 of ChatGPT's 698 occurrences survived into the paired repeat, 38.7% replaced, and 810 of Claude's 1,476, 45.1% replaced.

Ties are reported, not broken. A "top ten" includes every company reaching the tenth-highest count, so a ten can hold more than ten companies, and retention is counted on that membership — which is how nine of ChatGPT's ten and eight of Claude's are counted. Both tens in week 2 therefore hold eleven companies: in ChatGPT's, tenth place is shared by three companies at 9 recommendations — SmartBug Media, KlientBoost and Bop Design; in Claude's, by two at 14 — SmartBug Media and Belkins. Inside Claude's week-1 ten, Omniscient Digital and WebFX share 15.

The repeat pattern inside each week, counted per assistant. A company-and-question pair can only be counted here if that question was asked three times, so ChatGPT's columns cover the eighteen questions asked three times in both weeks, and Claude's cover all twenty.

How far apart the two assistants are: 57 companies were recommended by both in week 1 and 50 in week 2, and the overlap of the two sets fell from 0.238 to 0.225 on the Jaccard measure.

Site access, in full. The same 149 websites as in issue 1 were re-tested: what robots.txt declares, and separately how each site answered our requests under four named crawler user-agents, with every non-200 answer re-tested three times. Eighteen of our 596 named-agent observations answered differently than a week earlier. Two of them are the real refusals in the issue: stackmatix.com answering 403 to GPTBot and ClaudeBot on every attempt. Five moved the other way, to a normal answer — techcxo.com on three agents (504, 429 and 504 in week 1) and gorilla76.com on two. Ten are 200 answers that became 429, across seven sites: concurate.com, directiveconsulting.com, foundationinc.co, hingemarketing.com, nogood.io, schermer.co and velocitypartners.com. The eighteenth is smartsites.com, which gave us no answer at all on one agent where it answered last week. A 429 is a limit on how fast anyone may ask, including us, and it is never counted here as a block. Counted by site: 14 refuse persistently, 13 rate-limited at least one agent, 122 answered all four normally. On llms.txt, refinelabs.com and understoryagency.com added the file and foundationinc.co and smartsites.com no longer serve it; linkflow.ai stopped serving robots.txt altogether.

What each number is counted on. This measurement holds 118 answers: Claude answered all sixty of its asks, ChatGPT fifty-eight, and the two questions it answered twice are excluded from the table of repeat survival, where three repeats are the unit. Of site access, the first network answered on every one of the 149 websites, and the second network covers 201 of the 298 observations it could have taken this week against 241 of 298 in week 1 — a limit on how fast we may ask, not a property of the sites. The 97 observations not taken are recorded as not taken, never as a site answering normally, and where both networks answered they agree with each other on every site, including the refusal that is new this week.

Measured 26 October 2026, 06:01–06:33 UTC. Twenty questions frozen 19 October 2026, three repeats per assistant, 118 of 120 answers.