100 Questions, Twenty-first Category: AI Writing Tools, Where the Model's Own Product Is Picked Most and Specialists Win the Named Uses

ChatGPT leads the general questions

ChatGPT is named as an option in 58 of 100 answers, recommended in all 58 and picked as the answer's own choice in 50. Jasper is named in 43 and picked in 35, Grammarly 42 and 34, Claude 36 and 28. On the seven category questions that name no audience or use (best, best in 2026, free, cheapest, easiest, most popular, top rated), ChatGPT is picked in all seven. All 20 direct advice questions produced a pick: ChatGPT was among the picks in 13, Grammarly in 9, Jasper in 8, Claude in 5. Eight of the 100 questions name ChatGPT; in the other 92 it is named in 52 answers and picked in 45. It is rarely the only choice: in 47 of its 50 picks the answer picks another tool too.

Read that with the obvious caveat

ChatGPT is the system answering these questions, and these data cannot say whether its lead reflects the category or a preference for its own product. They can show what sits around it. ChatGPT's own site ranked in Google's top ten for none of the 100 questions. openai.com or chatgpt.com is cited in 15 answers, and ChatGPT is picked in 13 of them. No answer names ChatGPT as the system giving the answer. In nine more answers ChatGPT is named but not offered as an option: six times as an AI search surface whose visibility Writesonic, Surfer or Frase tracks, once as a point of comparison, once as the model inside another product and once only in an offer to compare tools later. The other general assistants trail: Claude is picked in 28 answers (ChatGPT is picked too in 24 of them), Microsoft Copilot in 7, Gemini in 4, Perplexity in 3. Both ChatGPT-versus-Claude questions pick both: one pairs ChatGPT with a versatile working partner and Claude with prose-first writing, the other says "I'd lean ChatGPT overall" for a novel and suggests testing both on your own voice. Asked for alternatives to ChatGPT, both answers put Claude first.

Specialists win the named uses

In all 15 questions that name a use or a platform, a specialist or the platform's own assistant is picked, and ChatGPT is picked beside it in 9. Sudowrite and Novelcrafter are picked in all four fiction and novel questions, ChatGPT in three of them. Shopify Magic is picked in all three product description and ecommerce questions. Surfer is picked in both SEO content questions: one answer picks it alone ("Surfer SEO is my pick"), the other pairs ChatGPT or Claude for drafting with Surfer or Frase for optimization. Lavender (with ChatGPT) is the pick for cold email, Grantable (with ChatGPT) for grant writing, QuillBot, Wordtune and Grammarly for paraphrasing, and Grammarly, ProWritingAid, LanguageTool and QuillBot for the grammar checker question. Microsoft Copilot is picked for Microsoft 365, and Gemini for Google Workspace, where ChatGPT appears only in an offer to compare tools later. Asked which tool is safest for confidential company documents, the answer picks Microsoft Copilot alone.

Almost every answer reaches a verdict

Every answer recommends at least one dictionary brand, and 96 of 100 pick one. The four without a pick give each option a fit and stop there: two three-way head-to-heads (Jasper versus Copy.ai versus Writesonic; ChatGPT versus Claude versus Gemini) and the alternatives questions for Writesonic and Sudowrite. Twenty-seven of 30 head-to-heads pick every brand named, each for a different case. The one that picks a side is Claude versus Gemini: Claude is "the one I'd test first", and Gemini is described as attractive for research-heavy writing but not picked. Being described is not always being picked: ChatGPT is recommended in 8 answers that do not pick it, Jasper and Grammarly in 7 each, and Gemini, recommended in 11, is picked in 4.

The widest reading list in the series

Across 100 answers there were 304 citation events to 143 domains, more distinct domains than in any earlier teardown. Vendors' own sites took 145 of them, 48 percent. Other third-party pages took 148, 49 percent, the highest share so far, and review-and-media sites 11. jasper.ai is the most-cited domain, in 25 answers, 18 of them to questions that never mention Jasper; grammarly.com is cited in 22 (15 not mentioning Grammarly) and openai.com in 14. The most-cited third party is dupple.com, in 12 answers across five roundup pages, and it ranked in Google's top ten for none of those 12 questions. Reddit ranks in Google's top ten for 93 of the 100 questions and is cited in none; Reddit and Wikipedia have zero citations, as in every teardown so far.

Of the 285 picked brand mentions, 240 went to brands whose own website did not rank in Google's top ten for the question, 84 percent. Google's top ten is 89 percent third-party pages. Forty-one of the 304 citation events involved a domain that also sat in Google's top ten for that query, 13 percent, level with SEO tools for the lowest overlap in the series.

Name the job you do better than a general assistant. In all 15 questions that name a use or a platform, a specialist or the platform's own assistant was picked. On the seven category questions that name no use, ChatGPT was picked in all seven. A specialist's pages should say plainly which job it is for.

Your own site is read for other people's questions. jasper.ai was cited in 18 answers to questions that do not mention Jasper, grammarly.com in 15. Use-case, pricing and comparison pages are material the model reaches for.

Roundup sites are part of your footprint. dupple.com was cited in 12 answers without ranking for any of them. If a roundup library covers your category, check whether you are in it and what it says.

Being described is not being picked. Gemini is recommended for a stated case in 11 answers and picked in 4. A fit label in a list of options is visibility, not a verdict.

One run per query means answer variance is unmeasured; Benchmark 01 runs each query three times across three surfaces. The teardowns are collected on different days, so differences between categories mix the category with the date. The brand dictionary covers the general AI writing tools market and deliberately excludes vertical tools, so their appearances are described in prose and not counted. The readers are AI models, not people: they follow a written protocol and agree with each other closely, but a shared blind spot would not show up as disagreement. The line between picked and recommended is a judgment, documented in the protocol with examples. Sentiment is not coded. No vendor-tool cross-check was read for this category. Google's control counts a brand as ranking only when its own domain is in the top ten; a listicle that features the brand does not count. One run per question, one day, United States, English, logged out. Answers vary between runs, so these are frequencies for this sample. The main limitation is structural: ChatGPT is judging a category that includes ChatGPT, and a single surface cannot separate the category's shape from a preference for its own product; that needs the same questions asked of other assistants. The ChatGPT row also counts OpenAI and GPT model names. The codes come from AI readers, not people; the two readers agreed on 97.3 percent of codes, and the calls settled by a third reader are listed in coder-notes.md. The line between picked and recommended is a judgment, and a different protocol would move some counts. Writer is matched in the dictionary only by its domain and model name; the readers added it where they saw it named, and some mentions may still be missed.

Every number above can be recomputed from these files. CC BY 4.0: use them, cite the page.

• queries.csv: the 100 queries with intent labels.
• brands.csv: the 53-brand dictionary with aliases and canonical domains.
• mentions.csv: 414 coded brand mentions with position, type, a 0/1 picked column and whether the brand's domain was in Google's top ten.
• citations.csv: 304 citation events with domain class and Google overlap.
• observations.csv: one row per query with brand, recommendation, citation and control counts.
• dictionary-changes.md: a ChatGPT alias change tried after collection and reverted; the dictionary used is the one fixed before collection
• coder-notes.md: coding method, reader agreement, the calls settled by a third reader, and dictionary notes
• codebook.md: the rulebook, with dated amendments through v2.0.
• reader-protocol.md: the written protocol both readers coded against (codebook v2.0).
• picked_stats.json: the picked-level figures.

The earlier teardowns: Teardown 01, project management, Teardown 02, crm, Teardown 03, email marketing, Teardown 04, help desk, Teardown 05, accounting, Teardown 06, payment processing, Teardown 07, payroll, Teardown 08, hr software, Teardown 09, applicant tracking, Teardown 10, password managers, Teardown 11, endpoint security, Teardown 12, business intelligence, Teardown 13, data warehouse and etl, Teardown 14, ecommerce platforms, Teardown 15, website builders, Teardown 16, scheduling, Teardown 17, video conferencing, Teardown 18, e-signature, Teardown 19, marketing automation, Teardown 20, seo tools.

AI writing tools is the last of the twenty categories planned for the series; payment processing was an extra. Here the model's picks track vendor pages and a set of roundup sites that do not rank. In password managers, 69 percent of citations went to the vendors' own websites. Finding out what the model is reading for your category is the first thing a Category Salience Brief does, with a query set you approve first.