Which AI engine recommends the most different tools?

GPT-OSS 120B named 115 distinct tools across 16 identical buying questions, the widest of the 8 engines measured. Llama 3.3 70B via Groq named 40, the narrowest. Both answered exactly the same questions, so the gap is a property of the engine rather than of the sample.

Does using live web search make an AI engine name more tools?

Not reliably. In this snapshot the 3 engines with live retrieval averaged 94 distinct tools and the 5 answering from model memory averaged 80. Retrieval changes which tools appear far more than how many.

If my tool is invisible to AI, which engine should I target first?

The one with the widest door. An engine that names 115 tools across 16 questions has room for a newcomer in a way an engine naming 40 does not. Breadth is not the same as traffic, so weigh it against where your buyers actually ask.

What does it mean when only one engine names a tool?

Across the 8 engines, 128 tools were named by exactly one of them on this shared question set. For a vendor that is a foothold rather than a win: you exist in one answer stream and are absent from the rest, so a single-engine visibility check would report success and a buyer on any other engine would never see you.

How was this measured?

Every engine was asked the same buying questions and every tool it named was recorded in order. Figures on this page use only the 16 questions that all 8 engines answered, because 32 further questions went to a subset and counting those would measure our sampling rather than the engines. Captured names are filtered to remove prose fragments; that filter rejected 2 percent of raw mentions.

5 of 8 AI Engines Cited No Sources at All

By Vincent Couey. Published August 27, 2026. Original research · 8 min read

Most GEO advice assumes there is a citation to earn. On most engines there is not. Across 288 recorded answers from 8 AI engines, 5 of them returned tool recommendations with zero sources attached, over 240 answers.

The 3 that do cite produced all 102 citations in the corpus, and they disagree with each other about how much to cite by a factor of 4.2.

Sources per answer, by engine

Citations attached per recorded answer

Gemini 2.5 Flash-Lite3.38

Perplexity2.19

ChatGPT0.81

Llama 3.3 70B via Groq0

Cohere Command-A0

Llama 4 Scout0

GPT-OSS 120B0

Qwen 3.60

Grey bars are engines that attached no source to any answer. They still named tools, confidently, with nothing to check and nothing to influence.

Why this breaks the standard GEO playbook

Where citation work lands

3 engines

Retrieval-backed. They fetch pages, name them, and you can earn a place in that set.

Roundups, comparisons, forum threads
Third-party pages more than your own
Measurable: the citation either appears or it does not

Where it does not

5 engines

Answering from model memory. No fetch, no source list, no surface to influence.

Nothing to be cited by
Changes only when the model does
Invisible to a citation-based audit

Both columns recommend software to buyers with the same confidence. Only the left one has a lever attached to it, and most published GEO advice is written as though every engine were the left column.

Even the engines that cite disagree about how much

Among the 3 retrieval-backed engines, citation density is not a shared standard. Gemini 2.5 Flash-Lite attached 3.38 sources per answer; ChatGPT attached 0.81. That is a 4.2x spread on the same 16 questions.

So "get cited" is not one target either. An engine that lists 3.38 sources per answer has room for a page that an engine listing 0.81 does not.

Which domains the citing engines reached for

All 45 citations in this corpus came from those 3 engines. The sample is small and is reported as such, so read the shape rather than the ranking.

Citations by domain, top 10

klaviyo.com6

zapier.com5

pcmag.com5

youtube.com3

zdnet.com2

crm.org2

emailvendorselection.com2

pendo.io2

technologyadvice.com2

ventureharbour.com2

45 citations across 20 domains. The leader appears 6 times, which is why this is a shape and not a ranking.

The mix is the familiar one: a vendor's own domain, an automation directory, a legacy tech publisher, a video platform, a forum. What matters here is not the order but that the whole list exists for fewer than half the engines a buyer might use.

What to do about the half you cannot cite your way into

First · split the target
Two problems, not one
Citation work reaches 3 of 8 engines here. Budget it as covering part of the field, not the field.
Memory engines
Presence, not placement
A model-memory engine names what its training corpus contains a lot of. That is a long, slow function of how widely a product is written about generally, and it does not respond to a single earned link.
Measure per engine
A citation audit sees half the board
An audit built on citations reports nothing for 5 engines, and nothing reads as fine. Check whether you are NAMED separately from whether you are CITED.
Do not average
The two are not commensurable
Blending a citation score across engines that cite and engines that cannot produces a number describing no engine at all.

What this does not show

Small citation sample102 source mentions is enough to establish that 5 engines attach none, and not enough to rank domains reliably. The domain ladder is shape, not league table.

Silence is not incapacityA model-memory engine may cite in other modes or products. This measures the answers we recorded, in the configuration we recorded them.

One snapshotEngines add and remove retrieval constantly. An engine silent here can become a citing engine with one release.

Not a quality claimCiting sources does not make an answer better, and a confident uncited answer is not automatically wrong. This is about whether a lever exists, not about accuracy.

Are you named on the engines that cite nothing?

The free AI Visibility check reports per engine, including the ones a citation audit cannot see. Given that 5 of 8 here attach no sources, per-engine is the only report that describes a real position.

Run my free AI Visibility check ›

📊 Want the citation-gap method? The capture schema, the per-engine source counting, and the audit design that separates NAMED from CITED. Delivered instantly.

Bottom line

5 of 8 engines in this corpus recommended software to buyers and showed their working to nobody. The industry's standard answer to AI visibility, earn the citations, is a real strategy that addresses 3 of them.

The useful question is not how to get cited. It is which engines can cite you at all, and what you are going to do about the ones that cannot.

Data: The AI Recommendation Audit (2026), CC-BY, DOI 10.5281/zenodo.20767878. 102 source mentions across 288 recorded answers. Companion: how the citing engines choose their sources covers WHICH pages get cited; this page covers how many engines cite at all. See also how widely each engine casts.