EX-06. AI Coding Assistants. AI visibility audit.

Cursor

Cursor leads AI Coding Assistants. In problem-first, the lane looks different.

Across 270 responses, Cursor appears in 71.5% of category-named queries. When a buyer describes a symptom instead of naming the category, the shape changes.

18/63

Cursor in problem-first

Cursor's presence in symptom-shaped queries.

Overall category presence: 193 of 270 (71.5%). In problem-first specifically, 39 of 63 responses in this bucket name no vendor at all (61.9%). That's empty inventory the leader could plausibly claim.

How AI answered a real buyer question in this category.

Query. Bucket 8. Problem-first. Excerpted from published audit.

I want to code faster but I don't want the AI to break things

That's the eternal tension with AI-assisted coding! Here are practical strategies to move fast without letting AI wreck your codebase:

## Set Up Guardrails Before You Start

**Version control discipline**

- Commit frequently (before *and* after AI changes) so you can always `git reset`

- Work on feature branches, never let AI touch main directly

- Review diffs carefully before committing — `git diff` is your friend

**Strong test coverage**

- Tests are your safety net. If AI breaks something, tests catch it immediately

- Ask the AI to write tests *first*, then implement (TDD-style)

- Run tests automatically on save or before every commit

## Constrain the AI's Scope

**Small, focused changes**

- Ask for one function or one file change at a time, not sweeping refactors

- "Modify only this function" is much safer than "clean up this module"

- Review each change before moving to the next

**Give it context, not free rein**

- Point it at specific files rather than letting it explore

- Share your conventions, types, and interfaces upfront

- Use tools with read-only defaults when possible (approve edits explicitly)

## Use the Right Modes

- **Plan mode / dry runs**: Have the AI describe what it *will* do before doing it

- **Ask before act**: Configure your tool to require approval for file writes, shell commands, and deletions

- **Sandbox risky work**: Use dev containers, VMs, or...

Vendors named in this response NONE

Appearance rank and pick-share rank tell different stories.

Every brand ranked two ways. Appearance = how often the brand is named at all. Pick share = how often the brand is the SOLE or PRIMARY pick, as a percentage of all 270 responses. The gap between the two is the recommendation gap.

#	Brand	Appear	App %	Pick %
01	Cursor	193	71.5%	30.7%
02	Copilot	182	67.4%	8.5%
03	GitHub Copilot	170	63.0%	8.5%
04	Claude Code	109	40.4%	9.3%
05	Windsurf	84	31.1%	3.3%
06	Codeium	64	23.7%	0.4%
07	Tabnine	54	20.0%	0.0%
08	Aider	48	17.8%	0.0%
09	Cline	38	14.1%	0.4%
10	Replit	34	12.6%	0.0%
11	Continue	33	12.2%	0.0%
12	Amazon Q Developer	30	11.1%	0.4%

Cursor appears in 71.5% of responses and is the sole or primary pick in 30.7%. That is a 40.8-point gap.

Top-mentioned brand in each type of buyer query.

Bucket	Top	n	Runner-up
Category discovery	Claude Code	36	Copilot 36
Comparison	Cursor	36	Claude Code 19
Segment fit	Copilot	32	Cursor 32
Evaluation	Copilot	28	GitHub Copilot 27
Pricing	Cursor	23	Copilot 16
Trust	Copilot	10	Cursor 9
Switching	Copilot	18	Cursor 18
Problem-first	Copilot	23	Cursor 18

Problem-first zoom.

The highest-leverage queries: the buyer describes a symptom rather than naming the category.

#	Brand	Count	%
01	Copilot	23	36.5%
02	Cursor	18	28.6%
03	Claude Code	15	23.8%
04	GitHub Copilot	14	22.2%
05	Windsurf	8	12.7%
n/a	Zero-vendor responses	39	61.9%

The domains the assistants pulled from.

Two lists. Vendor-owned and editorial sources on the left. SEO listing sites on the right. The split reveals which channel is doing the citation work.

Established, own content

forum.cursor.com Cursor32
cursor.com Cursor30
dev.to118
nxcode.io117
tech-insider.org112
qodo.ai62
verdent.ai59
zapier.com55
arxiv.org48
augmentcode.com47

SEO listing sites

none in this class0

Methodology and caveats.

Run date. 2026-07-24 UTC.
Sample. 30 buyer questions across 8 buckets, 3 samples per engine. Target 270 responses.
Engines. anthropic, openai, perplexity.
API is not the consumer app. Directionally representative, not literal.
Answers vary. Every prompt sampled 3 times per engine. Frequencies, not binaries.
Two-stage matcher (v0.4). Brand extraction is case-insensitive, word-boundary-anchored. Vendors whose canonical name overlaps with a common English word are flagged ambiguous in the config, and every candidate hit goes to a second-stage LLM adjudication (Claude Haiku 4.5) that reads the surrounding context and answers whether the token is the company or generic usage.
Recommendation classification (v0.5). Every confirmed brand mention runs through a second LLM pass (Claude Haiku 4.5) that classifies the treatment of that brand in that response as one of six codes: SOLE (only vendor named as the answer), PRIMARY (named as the top pick), ALTERNATIVE (named in a short recommended set), CONDITIONAL (recommended only for a stated narrow case), NEUTRAL (mentioned without endorsement, or cited as a source only), or UNFAVORABLE (named with a negative or dismissive frame). Pick share of total = (SOLE + PRIMARY) ÷ total responses.
No editorializing. This exhibit reads the data. It does not claim causation.
Right of reply. Any named brand can request their own free audit. Findings publish alongside.

Want an audit like this for your brand?

Every 30-day pilot starts with one. Fixed fee. Ad spend billed to your account. Written verdict on day 30.

Apply for a pilot

Or read the rest of the research library.