Machine Relations

ChatGPT Recommends Your Competitors and Not You: Where It Drops You and What Changes It

Engines cited LinkedIn in 181, Reddit in 165 and YouTube in 125 of 1,951 B2B problem-first runs. Find where ChatGPT drops you and what to check first.

Jaxon ParrottJan 28, 2026

When ChatGPT names your competitors and not you, you have dropped out at one of three stages: no source it retrieved names you, a source did but the answer left you out, or the answer named you without recommending you. One run cannot tell which. Check each engine over repeated runs, then check whether any independent source names you.

That makes a citation gap a sourcing question first: the studies below tie being named to what independent sources say. The Machine Relations Index (MRI), published by Machine Relations Research, shows which sources engines cite. It counts cited domains, not whether an answer names a brand, so why one company is skipped is outside what it records. For problem-first questions in 11 B2B categories it shows where engines look: LinkedIn, Reddit and YouTube first, and a different set of domains in every category.

Key Takeaways

One answer is not a diagnosis. In the Index question panel, two single runs of the same question shared 14.3% of their cited volume: 31.6% on the same engine, 9.0% across engines.
Find your stage. A 37,000-run audit separates "not retrieved", "retrieved but not mentioned" and "mentioned but not recommended". Each points to different work.
Problem questions cite fewer, different sources. Across 1,951 problem-first runs in 11 B2B categories, engines cited 4.85 domains per run, against 7.30 for best-tools questions; LinkedIn (9.3% of runs), Reddit (8.5%) and YouTube (6.4%) led.
The cited set belongs to your category. arXiv and Reddit lead AI Infrastructure; GitHub leads Cybersecurity; niche vendor sites lead Industrial and Logistics & Freight.
A citation is not a recommendation. In one Google AI Overviews study, a brand's own list page was cited but the brand went unrecommended in 224 of 323 cases.
What changes it is independent sources that name you. The evidence is correlational or vendor-run, so measure repeated runs before and after any spend.

Why does ChatGPT name my competitors and not me?

The evidence supports three drop-out stages, and each points to different work: retrieval, extraction or recommendation. An arXiv preprint by authors affiliated with Unusual, a company that sells AI-visibility services, audited about 37,000 runs across four OpenAI and Anthropic production model configurations, 215 commercial prompts and a 533-brand catalog. It classed each brand in each run by the last stage reached, and states that the labels describe where a brand dropped out, not why.

Stage	What you see	What the audit found	First check
1. Not retrieved	No source shown for any answer names you	48 to 52% of the two smallest brand tiers never surfaced in any of the 37,000 runs	Crawler access, then whether any independent page names you
2. Retrieved, not mentioned	A source the engine shows names you; the answer does not	Causes not isolated: relevance, list length and buyer fit are named as possibilities	Read that source: is the claim about you specific?
3. Mentioned, not recommended	The answer lists you as an alternative	Category leaders won 25 to 41% of the recommendation slots they reached	Who else recommends you, and in what words

Engines do not always search. OpenAI says ChatGPT may search the web automatically when a question would benefit from current information, and Anthropic says the same of Claude, so a given answer may carry no sources. Coverage also differs: in the Index release, Gemini returned a citation in 2,153 of 6,384 returned runs.

Is one ChatGPT answer enough to diagnose it?

No. In the Index question panel, two single runs of the same question shared 14.3% of their cited volume, and SparkToro found under a 1 in 100 chance that two responses from ChatGPT or Google's AI carry the same list of brands. The Machine Relations Research study compared halves of 980 questions across 22,620 runs from July 6 to October 20, 2026.

Comparison	Shared cited volume	Share of rerun level kept
Other half of the same question (a rerun)	46.1%	100%
Half of a reworded question	39.4%	77 to 88%
Half of a broader or narrower question	23.0%	44 to 49%
Two single runs of one question	14.3%	not reported
Two single runs on the same engine	31.6%	not reported
Two single runs on different engines	9.0%	not reported

That study counts cited domains, not brand mentions, so it shows that the sources behind an answer shift, not that competitor names do. SparkToro's January 2026 study measured names: 600 volunteers ran 12 prompts through ChatGPT, Claude and Google's AI 2,961 times, and SparkToro recommended 60 to 100 runs per prompt before reading a list (it disclosed that its research partner works at an AI-tracking startup). The arXiv preprint "Don't Measure Once" reaches the same conclusion: treat visibility as a distribution. In the Index panel, a domain cited in one run was cited again in the other half 37.8% of the time, and 95.6% when it had appeared in five or more runs.

What do engines cite when a buyer states the problem?

In the 11 B2B categories whose problem-first segment clears the Index evidence floor, engines cited LinkedIn in 181 of 1,951 runs (9.3%), Reddit in 165 (8.5%) and YouTube in 125 (6.4%); the most-cited classified editorial publication, TechRadar, appeared in 64 (3.3%). A problem-first question is one where the buyer describes a problem rather than naming a tool. The table compares it with the five other question shapes.

Question shape	Runs (published segments)	Cited domains per run	Editorial publications (floor)	Community and social (floor)	Vendor-owned (floor)	Other observed source (ceiling)
Best tools	1,844	7.30	4.8%	4.0%	12.0%	68.1%
How buyers choose	1,855	5.84	4.1%	4.6%	10.6%	72.0%
Is it worth it	1,965	5.86	4.4%	4.3%	12.6%	70.9%
Problem-first research	1,951	4.85	5.6%	6.5%	10.6%	68.9%
Top lists	1,785	7.39	4.0%	2.9%	10.7%	72.8%
Comparisons	1,820	5.54	4.8%	5.6%	13.5%	63.9%

Shares are of domain-run citations. The Index classes a minority of sources, so each class share is a floor and "Other observed source" is a ceiling; gaps of a point or two between shapes sit inside what unclassified sources could hide. The stable reading is the third column: a problem-first answer cites 4.85 domains, against 7.30 for best-tools questions, so each source carries more weight. YouTube is classed as a search or media platform (1.4% of problem-first citations), not as community. The leaders differ by category:

Category (problem-first runs)	Three most-cited domains (runs, share)	Editorial share (floor)
AI Infrastructure (225)	arxiv.org 30 (13.3%), reddit.com 28 (12.4%), dev.to 25 (11.1%)	5.5%
AI Security & Privacy (224)	techradar.com 31 (13.8%), arxiv.org 30 (13.4%), conferbot.com 25 (11.2%)	17.5%
AI Visibility & GEO (249)	youtube.com 46 (18.5%), reddit.com 42 (16.9%), linkedin.com 39 (15.7%)	5.1%
Cybersecurity (205)	github.com 37 (18.0%), nhimg.org 27 (13.2%), microsoft.com 24 (11.7%)	3.6%
Enterprise Software (178)	reddit.com 16 (9.0%), zapier.com 16 (9.0%), cloudeagle.ai 13 (7.3%)	4.3%
Fintech (183)	chargeflow.io 31 (16.9%), rexi.finance 19 (10.4%), stripe.com 19 (10.4%)	3.9%
HR & Talent (169)	linkedin.com 27 (16.0%), paychex.com 25 (14.8%), paycom.com 20 (11.8%)	4.1%
Industrial (133)	fabrico.io 19 (14.3%), customconveyor.com 16 (12.0%), oxmaint.com 14 (10.5%)	0.5%
Logistics & Freight (100)	dashdoc.com 16 (16.0%), opendock.com 12 (12.0%), withvector.com 11 (11.0%)	2.4%
Martech & Advertising (101)	cometly.com 13 (12.9%), braze.com 9 (8.9%), deployteq.com 7 (6.9%)	4.2%
Deep Tech & Hardware (184)	analog.com 35 (19.0%), ti.com 25 (13.6%), arduino.cc 18 (9.8%)	1.7%

In the AI Visibility & GEO row, Semrush (31 runs, 12.4%) is the fourth source and the highest classed vendor-owned; most of the other leading domains are classed "Other observed source", so this page does not split them. Pooled rates also lean toward engines that cite more: in the same Index question panel, Perplexity supplied 50.7% of domain-run pairs.

What should I check first, and in what order?

Freeze the wording, run it repeatedly on every engine your buyers use, rule out crawler blocks, find your stage, tally the source classes shown, then test whether any independent source names you. This is a screen, not a verdict; the full method is in How to Run an AI Citation Gap Analysis in Five Steps.

Freeze three wordings of the buyer's question, without your brand name: the original, a reworded version and a narrower one. A reworded question kept 77 to 88% of the rerun overlap in the Index study; a narrower one is a different question to track on its own.
Run each wording at least three times on each engine, with search on where offered, and log the brands named, the sources shown and whether you appear. Three runs only screen; SparkToro's 60 to 100 is the standard for reading a list.
Rule out blocks. OpenAI's crawler documentation says sites opted out of OAI-SearchBot are not shown in ChatGPT search answers, and Perplexity's recommends allowing PerplexityBot and, behind a firewall, its published IP ranges. Google's AI features page says a page must be indexed and eligible to show with a snippet, with no additional technical requirements. A block can explain a skip; an open door explains nothing.
Assign the stage per engine from the first table, using only the sources the engine shows.
Tally the source classes shown where competitors are named: vendor sites, communities and video, editorial and trade publications, review and market databases, research. Compare with your category's problem-first page above and its best-tools and top-list pages.
Search the cited domains, and your category's leaders, for your name. If no independent page names you, the gap is sourcing.

What you find	What it means	Fix first
A crawler block or a noindex	The engine may never fetch you	Remove it; OpenAI says a robots.txt change takes about 24 hours to apply
Open to crawlers, and no independent page names you	Sourcing gap (stage 1)	Earn coverage in the source classes your category's answers cite
Independent pages name you, but the engine shows other sources	You are not yet among your category's leading sources	Read which class wins in your category and get named there
A shown source names you, the answer does not	Extraction or specificity (stage 2)	Make the claim about you specific and quotable in that source
Your own page is cited, a competitor is recommended	Citation is not recommendation (stage 3)	Stop ranking yourself first in your own lists; build third-party recommendations
The engine shows no sources	The answer may come from memory	Re-run with search on and judge over repeated runs

What actually changes it?

Being named by independent sources the engines already retrieve for your category; the evidence is correlational or vendor-run, so measure repeated runs before and after any spend. Five findings sit behind that.

Evidence	What it supports	Limit
Ahrefs, 75,000 brands: web mentions correlate 0.664 with AI Overview mentions, backlinks 0.218	Off-site presence tracks AI visibility more than links do	Spearman correlation, AI Overviews only; Ahrefs states correlation is not causation
Chen et al., University of Toronto: AI search favors earned over brand-owned and social sources	Source type matters, and niche brands face a big-brand bias	Unrefereed arXiv preprint
Aggarwal et al., KDD 2024: content tactics lifted visibility by up to 40%	On-page work helps pages that are retrieved	Benchmark queries; effect varies by domain
Stacker and Scrunch, March 2026: 87 stories, 2,600+ prompts, 8 platforms, median 239% lift in citations	Third-party distribution raised citations to those stories	Run by Stacker, which sells the distribution it measured; counts citations, not brand naming
Lily Ray's analysis of 100 B2B queries, 80 of which triggered an AI Overview: own-brand lists cited 323 times, brand unrecommended in 224	Citation and recommendation separate	Google AI Overviews, B2B software, April to June 2026

Google adds a caution. Its optimization guide says llms.txt files, special markup and rewriting content only for AI systems are not needed, and that seeking inauthentic mentions "isn't as helpful as it might seem", while noting its generative features can show what is said about products across blogs, videos and forums. The sources worth earning are real editorial and community coverage, not manufactured mentions.

Disclosure: AuthorityTech is an earned-media agency that secures coverage for clients in the outlets engines already cite, so it has an interest in how this page reads. Its approach is to read the cited sources for the client's category in the Index, secure placements in those outlets, then re-run the buyer's questions across engines and report what changed. No placement is promised to be cited: the Index records what engines cited, not what a placement caused. AuthorityTech publishes no rate card; scope and fees are set on a strategy call, and placements are paid for after publication, with funds held in escrow until a placement is live. AuthorityTech is The World's First Machine Relations Agency. Start your visibility audit →

What the Index cannot tell you

Why one brand is skipped is outside what it records. It counts cited domains per run, never whether an answer names a company, so a stage-2 or stage-3 miss does not appear in it.
It does not show your standing. It runs the market's questions, not yours; your own prompts, run repeatedly, are the measure.
Engines in the panel are API models. The manifest lists gpt-4.1-mini for ChatGPT, claude-haiku-4-5 for Claude and sonar for Perplexity, which can differ from consumer apps, and marks Gemini coverage as degraded.
The overlap figures need the Index run ledger. Machine Relations Research states they cannot be recomputed from the public file alone.

Method note. Index figures are the public artifact for release mri_score_v2.1+2026-10-20+bf82d9d98978 (window May 10 to October 20, 2026; 28,201 runs; six engines; 55,217,722 bytes, sha256 beginning bf82d9d98978). They sum each domain's per-segment rows (runs cited) for the six question shapes across the 14 B2B categories, keeping only segments with status published: AI Infrastructure, AI Security & Privacy, AI Visibility & GEO, Cybersecurity, Enterprise Software, Fintech, HR & Talent, Industrial, Logistics & Freight, Martech & Advertising, Professional Services, Sales, GTM & CRM, VC & Private Equity and Deep Tech & Hardware. Problem-first segments in Professional Services, Sales, GTM & CRM and VC & Private Equity (65 of 2,016 B2B problem-first runs) are below the evidence floor of 10 runs on 7 dates and are excluded, leaving 11 categories and 1,951 runs; the other shapes use their own published segments, so their run counts differ (1,785 to 1,965). A run is one question on one engine on one day, and a domain counts once per run. Cited domains per run is domain-run citations divided by observed runs, including runs that cited nothing: 9,458 over 1,951 for problem-first, 13,462 over 1,844 for best tools. Source classes are the Index's source_role labels. Domain shares are runs citing the domain over runs observed in the segments summed, a ranking this page makes and the Index does not publish; ties are ordered as the Index ranks them. Re-query the public JSON artifact, the release manifest, the category pages above and the domain pages for linkedin.com, reddit.com and youtube.com. Perplexity's 50.7% share and every overlap figure come from the Machine Relations Research study linked above.

Frequently asked questions

Why does ChatGPT recommend my competitors and not my company?

Three drop-out points are documented: no retrieved source names you, a retrieved source names you but the answer leaves you out, or the answer names you without recommending you. In a 37,000-run audit, 48 to 52% of the two smallest brand tiers never surfaced in any run. Why a given company is skipped is outside what the Index records.

What should I check first when ChatGPT names competitors and not me?

Freeze three wordings of the buyer's question, run each at least three times on every engine your buyers use, and rule out crawler blocks. Then find your stage and tally the source classes the engine shows. That is a screen; SparkToro recommended 60 to 100 runs per prompt.

Does one ChatGPT answer prove my brand is invisible?

No. In the Index question panel, two single runs of one question shared 14.3% of their cited volume, and SparkToro found under a 1 in 100 chance of the same brand list in two responses.

Which sources do AI engines cite when a buyer describes a problem?

In 11 B2B categories, LinkedIn was cited in 181 of 1,951 problem-first runs, Reddit in 165 and YouTube in 125, with 4.85 cited domains per run against 7.30 for best-tools questions. The leaders differ by category.

Will publishing more on my own site fix it?

Own pages make you eligible and easy to quote, and content tactics lifted visibility by up to 40% in one benchmark. They do not create independent sources, and a company's own list page can be cited while a competitor is recommended: 224 of 323 cases in one Google AI Overviews study. The owned-site question has its own page: Why AI Search Engines Ignore Your Website.

Can a placement fix a citation gap?

A placement can add an independent source that names you. The Index records what engines cited, not what a placement caused, so AuthorityTech reports what engines cite before and after a placement and does not promise that any placement will be cited.

What is the Machine Relations Index?

It is the benchmark of which sources AI answer engines cite, published by Machine Relations Research at machinerelations.ai and updated daily. Machine Relations is the discipline Jaxon Parrott coined in 2024; AuthorityTech, which he founded in 2018, is The World's First Machine Relations Agency. The Index is Machine Relations Research's instrument, not an AuthorityTech tool.