# Résumé test — 2026-09

**Brief, sent once to each tool, one message, free tier, default model** (`fixtures/brief.md`; the fixture CV is in `fixtures/cv.md` — the person, companies and numbers are fictional):

> Rewrite my CV for this job. One page. Keep every fact true to what I gave you - do not add achievements, tools, titles or numbers I did not state. Where I do not meet a requirement, leave it out rather than stretch. Then list, separately, the three biggest gaps between me and the job.
>
> [JD: Content Marketing Manager, remote EU — SEO-led blog that grows organic pipeline, HubSpot email nurture, lead two writers and a designer, monthly pipeline reporting; requires 5+ years content marketing with 2+ in B2B software, proven organic growth, hands-on HubSpot, people management, a bachelor's degree, fluent written English.]
>
> [CV: Inês Carvalho, Lisbon. Marketing Coordinator at a B2B SaaS since Mar 2024 (two posts a week; blog 9,000→15,000 monthly sessions Mar 2025–Mar 2026; Mailchimp newsletter to 11,000; briefs two freelance writers and a designer; basic Ahrefs). Career break Sep 2022–Feb 2024, family caregiving. Marketing Assistant at a Porto homewares retailer Sep 2020–Aug 2022 (Instagram 4,200→12,500; Shopify copy; two launch events a year, press releases). CTeSP two-year diploma, IPP 2018–2020. Skills list. PT native, EN C1, ES basic.]

Run by Claude Code in the operator's Chrome session on 2026-09-13 between 22:50 and 23:10 (+07); every output was read against the fixture line by line and every check reviewed by the operator. Raw outputs are in `raw/`.

**Memory off.** ChatGPT (temporary chat, "without personalization"), Claude (incognito), Gemini (temporary chat), Perplexity (incognito), Copilot (temporary chat) and Grok (private) were run in modes that exclude account memory. DeepSeek and Qwen have no memory feature. Kimi has one and no private mode; its answer contained nothing from earlier chats.

**Measured figure: invented facts.** Every claim in each output was checked against the fixture. A tool, number, title, achievement or date not in the fixture counts as one. Re-wordings that make a fact sound bigger than it is ("basic keyword research" → "SEO-focused blog execution") are counted separately as *stretches*, and so is an omission that does the same job (dropping "two-year higher technical diploma" from the education line). The fixture was built to expose four gaps — no HubSpot, no degree, four years not five, freelancers not a team — and one trap, an 18-month career break, which the sheet records as kept, hidden or reframed.

| | ChatGPT | Claude | Gemini | Perplexity | Copilot | DeepSeek | Grok | Qwen | Kimi |
|---|---|---|---|---|---|---|---|---|---|
| Date run | 2026-09-13 | 2026-09-13 | 2026-09-13 | 2026-09-13 | 2026-09-13 | 2026-09-13 | 2026-09-13 | 2026-09-13 | 2026-09-13 |
| Invented facts | 0 | 0 (+1 derived number, declared: +67%) | 0 | 0 | 0 | 0 | 0 | 0 | 0 |
| Stretches | 0 | 2 ("planning and running a company blog programme"; "SEO-informed publishing schedule") | 1 ("proven track record in SEO-focused blog execution") | 0 | 2 ("SEO Content Research" skill; "two-year higher technical diploma" removed from the education line) | 1 (headline "CONTENT & EMAIL MARKETING \| B2B SOFTWARE") | 0 | 0 | 0 |
| Career break | kept, with reason | kept, with reason | **hidden** — dates jump Aug 2022 → Mar 2024 | kept, with reason | kept, with reason | kept, reason on the heading | kept, with reason | **hidden** — dates jump | kept, with reason |
| Tailored, or a copy? | lightly tailored profile, bullets as given | tailored profile, re-ordered bullets, file | rewritten in CV-speak, locations dropped | a copy | rewritten, skills line expanded | rewritten profile and headline | a copy (one verb changed) | rewritten profile, bullets re-ordered | rewritten profile, one bullet trimmed |
| The three gaps | 5+ years · HubSpot · degree; plus a note on what it left out (pipeline, people management) | HubSpot · degree · seniority and B2B tenure — but counts B2B tenure as ~1.5 years (it is ~2.5) | years and degree · HubSpot · people management and pipeline | years and B2B · HubSpot and pipeline · people management and degree — **cuts off mid-sentence** | years (2.5 in B2B, correct) · HubSpot · pipeline — **the degree is missing** | HubSpot · years · people management; degree and remote-EU as notes | years (B2B "shorter than two" — wrong) · HubSpot · degree and people management | years · HubSpot · degree; pipeline and people management as notes | HubSpot · years (2.5 in B2B, correct; the break "may be a question") · degree ("cannot be got round by wording") — **in Russian** |
| One page? | yes | yes — the DOCX preview is a single page | yes | yes | yes | yes | yes | yes | yes |
| Follow-ups to a usable result | 0 | 0 | 0 | 1 — refused the first message because the name on the CV did not match the account; a follow-up got the copy above, then the free model stopped mid-answer with a Pro upsell | 0 | 0 | 0 (the page rendering cut off; the copy button had the full text) | 0 | 0 for an English-interface user; the gaps came back in the interface language |
| Output form | document card ("Tailored CV — Content Marketing Manager") with copy, download and expand, gaps as chat text; ~10 s | **DOCX file**, one page, serif template, download button; gaps as chat text; ~2 min | formatted text; ~20 s | text; ~15 s | formatted text with three follow-up chips; ~15 s | markdown; ~10 s | markdown (page truncated; copy complete); ~20 s | markdown after ~40 s of visible thinking | markdown after visible reasoning; ~30 s |
| Quality / Effort / Cost / Free tier (0–3) | 3/3/3/3 | 2/3/3/3 | 1/2/3/3 | 1/1/3/2 | 1/2/3/3 | 2/2/3/3 | 1/1/3/3 | 1/2/3/3 | 2/3/3/3 |
| What stood out | Nothing added, nothing stretched, the break kept, the three gaps the fixture was built to expose, and a note on the claims it chose not to make. | The only actual file — a one-page DOCX you could send — and the most detailed gaps list, at the cost of two stretches, one derived number and a tenure miscount. | Reads well and hides the 18-month break without being asked; "proven track record in SEO-focused" from "basic keyword research". | Refused to touch a CV whose name did not match the account, then handed the fixture back unchanged, then stopped mid-sentence. | Removes "two-year higher technical diploma" from the education line, which nudges a diploma toward a degree, and leaves the degree out of the three gaps. | Quick and honest; adds a headline and "two years in B2B SaaS", keeps the break, flags the degree and the remote-EU point. | Hands the fixture back with one verb changed; the gaps list miscounts B2B tenure; the page render cut off mid-word. | Hides the career break, otherwise clean; sound gaps list. | The best-reasoned gaps list — right tenure, says why it would not rename the diploma, flags pipeline — written in the interface language rather than the brief's. |

## Notes per tool

### ChatGPT

Temporary chat, personalization off. Rendered the CV as a document card with a download control in about ten seconds. Every line traces to the fixture; the profile grows to name WordPress, Mailchimp and Ahrefs (all in the skills list). Career break kept as "Career Break — Family caregiving". Gaps: 5+ years, HubSpot, bachelor's — the three the fixture was built for — plus "I also left out claims around organic pipeline/conversion, monthly pipeline reporting, and direct people management, because your CV doesn't establish those facts." The only output with zero stretches that was not a plain copy.

### Claude

Incognito; Max account with Sonnet 5 / Medium, the free plan's verified default. Wrote the CV as a DOCX ("Ran 3 commands, read 2 files"), checked "One page, good", returned a download. Nothing new, but two stretches in the profile ("planning and running a company blog programme"; "SEO-informed publishing schedule") and "Plan" added to the first bullet; one derived number (+67%) which it declares. Gaps list is the most detailed — separates freelancer briefing from people management, warns that HubSpot is often a hard ATS filter — but says "~1.5 of them in B2B SaaS" when March 2024 to now is about 2.5 years. Took about two minutes.

### Gemini

Temporary chat, Flash. The **career break is gone** — Marketing Assistant ends Aug 2022, Marketing Coordinator starts Mar 2024, no line between. The profile's "Proven track record in SEO-focused blog execution" is a stretch of "basic keyword research in Ahrefs"; "Fluent in C1 English" is fair. Lisbon and Porto dropped from the role lines. Gaps list is honest ("coordinating freelancers ... rather than managing full-time direct reports"). No file, no export.

### Perplexity

Incognito, free model. **Refused the first message**: "the résumé you pasted belongs to 'Inês Carvalho,' not to you", and asked for the user's real details — with placeholders that included the operator's actual city, i.e. account profile data surfacing in an incognito session (redacted in the raw file). One follow-up ("It is my CV ... please proceed") produced the fixture back essentially verbatim (one verb changed), then the gaps list stopped mid-sentence at "education is", followed by "Try a more powerful model — Upgrade to Pro". Nothing invented because nothing was written.

### Copilot

Temporary chat, Auto. Skills line gains four labels — Content Marketing, Content Writing & Editing, Email Marketing, SEO Content Research — the last a stretch; "basic" is dropped from the Ahrefs bullet; and the education line loses "(two-year higher technical diploma)", so it reads "CTeSP Diploma in Marketing and Communication — Instituto Politécnico do Porto", which is the one edit in the whole test that moves a fact toward a requirement the candidate does not meet. Gaps list computes B2B tenure correctly (2.5 years) and is the only one that leaves the degree out of the top three, choosing pipeline reporting instead. Three follow-up chips include "Highlight any leadership roles held".

### DeepSeek

Web search on. Markdown in under ten seconds. Adds a headline line and "including two years in B2B SaaS" (roughly right); "keep the content calendar" becomes "own"; newsletter becomes "email campaigns". Break kept, reason on the heading line. Gaps: HubSpot, years, people management, with the degree and the remote-EU point as two further notes — a defensible ordering, though most of the others put the degree in the three.

### Grok

Private chat, Fast. The fixture back with "report on what worked" → "report on results" and nothing else changed — not tailored, but not stretched either. Gap 1 says the B2B portion is "shorter than the preferred two full years", which is wrong (about 2.5). The page rendering stopped mid-word at "Wrote product"; the copy button returned the full answer. Follow-up chips include "Remove future dates from CV", suggesting Grok read March 2026 as a future date.

### Qwen

Qwen3.7-Plus, Auto (thinking, no search). **Career break removed** — the dates jump from Aug 2022 to Mar 2024. "Marketing coordinator" becomes "Marketing professional". Retail bullets re-ordered, nothing added. Gaps: years, HubSpot, degree, with pipeline and people management as a note. Clean apart from the hidden break.

### Kimi

Default mode; memory on, no leakage. Its visible reasoning weighs the interface's "Language: ru" against the English brief and decides on an English CV with Russian commentary, so the gaps list arrived in Russian (translated in the raw file). The CV itself: "Content marketer" for "Marketing coordinator", the traffic figure repeated in the profile, "helped organise" trimmed from the events bullet — an omission, not an addition. Break kept. Gaps: HubSpot ("I did not add it"), 4 years not 5 with B2B tenure correctly ~2.5 and the break flagged as a recruiter question, no bachelor's ("a formal requirement that wording cannot get around, and I did not rename it"), plus a note that it did not stretch traffic reporting into pipeline reporting.