100 Help Desk Questions: What ChatGPT Recommends

The help desk software ChatGPT recommended most often was Zendesk, in 69 of 100 answers, and Zendesk was its own pick in 61 of the 100. In short: recommended nearly every time, picked most of the time.

Zendesk leads, and the field is deep

Zendesk leads a deep field: it appears in 70 answers, is recommended in 69 and picked in 61. Freshdesk appears in 60, is recommended in 59 and picked in 51; Help Scout 50, 49 and 45; Intercom 42, 42 and 36. That is a leader with a clear gap, as in project management, but with a longer competitive tail: Zoho Desk, Jira Service Management, HubSpot Service Hub and Gorgias each appear in 24 to 31 answers, are recommended in every one of them and picked in 20 to 24. In this category the gap between appearing and being recommended has all but closed; the gap that is left is between recommended and picked. Category queries name 6.1 brands per answer on average, level with project management for the most of the first four categories, and recommend 5.5 of them and pick 4.5. Seventy-two answers recommend three or more brands and 64 pick three or more.

Intercom's Fin AI agent is a special case: it is named in 18 answers, 16 of them in passing, typically as a feature of Intercom, and recommended as a product in its own right in 1. Freshworks' Freddy AI shows the same pattern, named in 10 and recommended in none. The AI agent is what the model talks about; the help desk is what it recommends.

When asked for a pick, Help Scout and Zendesk tie, and Help Scout owns the small end

On direct requests, Help Scout is picked as often as Zendesk, and more often for small teams. All 20 recommendation-intent queries recommended a dictionary brand, and all 20 picked at least one. Help Scout and Zendesk were each among the picks in 14 of them, Freshdesk in 12, Intercom in 10, Zoho Desk and Jira Service Management in 5 each. No query produced a single pick: every answer is a fork of two to six brands, Help Scout for the smaller team, Zendesk or Jira Service Management for scale, Gorgias for the store, Freshservice for the IT department. The small end is where the lanes show. Help Scout is picked for the 5-person support team, the solo founder, the small business owner who wants something simple and the team replacing a shared Gmail inbox; Zendesk is picked for only the first of those four.

Twenty-nine forks and one non-answer

All but one comparison answer is a fork. In 29 of 30 head-to-head queries ChatGPT recommended every brand named in the query, and in the same 29 it picked every one of them, each for its own case. Freshservice versus ServiceNow, a single winner in the original version of this page, is one of those forks: Freshservice for simplicity and speed to deploy, ServiceNow for enterprise depth. HelpDesk versus LiveChat produced no recommendation at all: the answer read it as a question about help desk software versus live chat in general.

The vertical prompts did not swap the field

No vertical prompt pushed the general field out. Schools, healthcare, law firms, a real estate brokerage, MSPs, IT teams: every one of those prompts still produced a recommendation and a pick from the general dictionary. Help desk joins email marketing as a category where a vertical word did not erase the general brands; in CRM and project management it did. Help desk software appears to be read by the model as horizontal in a way that CRM and project management are not.

Nearly four in five recommendations go to brands that do not rank

Of the 445 recommended brand mentions, 350 were for brands whose own website does not rank in Google's top ten organic results for that query, 79 percent. Of the 380 picked mentions, 291 were, 77 percent. Google's top ten for these queries is 71 percent third-party pages. Between the four categories the recommended figure now runs 70, 75, 79 and 85 percent (CRM, email marketing, help desk, project management).

The third most-cited source is a site called Macha

Macha takes third place among the most-cited sources. Across 100 answers there were 345 citation events to 113 domains, one per cited domain per answer. Zendesk.com leads with 32 and intercom.com follows with 21. Third is getmacha.com, cited 20 times, ahead of freshworks.com (18), G2 (15), helpscout.com (12) and gorgias.com (12). Macha is a vendor blog whose comparison posts (best free Zendesk alternatives, Freshdesk versus Help Scout versus Zendesk, cheaper Intercom alternatives) the model cited in 20 answers across all four intents. Behind it are commsadvisor.com (7), getomnichannel.com (6), interobservers.com (5), stackbriefly.com (5) and toolradar.com (5). Forty-six percent of citation events went to third-party sites that are neither review platforms nor recognized publishers, tied with CRM for the highest share of the first four categories, and only 44 percent went to vendors' own sites. Reddit and Wikipedia have zero citations, for the fourth teardown running.

Eighty of the 345 citation events involved a domain that also sat in Google's top ten for that query, 23 percent.

The table sets ChatGPT's help desk software answers beside the other categories tested earlier in the series, side by side under the same rulebook.

What ChatGPT's recommendations mean for help desk software brands comes down to four things, in order.

Find out who the model is actually reading. A vendor with a comparison-article program is the third most-cited source in this category, ahead of Freshworks and G2. That source is where the recommendation evidence comes from, and it is not on anyone's link-building list.

Own a lane and say it plainly. Help Scout owns "small team, email support" in the model's words and is picked in all four small-scale recommendation prompts, where Zendesk is picked in one. Zendesk owns scale. Nearly every brand the model offers here is recommended for something; what separates them is whether the answer makes them its pick for the buyer in front of it.

Treat the AI agent as the conversation and the help desk as the sale. Fin is named in 18 answers and Freddy in 10, and between them they are recommended once. Product marketing that leads with the agent is feeding the model description, not selection.

Do not assume vertical prompts erase you. In this category they did not. Measure it for yours before you build vertical pages you may not need.

The limitations of the ChatGPT help desk software test start with variance: 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 help desk software 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. Fin, Freddy AI and Zendesk AI are sub-products coded as separate brands so that their appearances could be counted; their parents' figures do not include them. The dictionary excludes vertical tools by design.

Every number above can be recomputed from the ChatGPT help desk software dataset below, so download it and check it rather than believe it. CC BY 4.0: use the files, cite the page.

• queries.csv: the 100 queries with intent labels.
• brands.csv: the 56-brand dictionary with aliases and canonical domains.
• mentions.csv: 531 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: 345 citation events with domain class and Google overlap.
• observations.csv: one row per query with brand, recommendation, citation and control counts.
• 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. This teardown is also being published on the High Salience Substack.

The ChatGPT test we ran on help desk software can be run on your own category too. In the wider series: four categories, four different shapes.

A favorite, a big four, a seven-way spread, and a deep field with an unknown source feeding it. The only way to know which shape your category has is to measure it. A Category Salience Brief runs this exact method on it, with a query set you approve first.

For software companies, data like this is where our work as a SaaS SEO agency for AI search begins.