Onboarding UX examples: what 95 SaaS products ask after sign-up
We read 304 onboarding screens from 95 SaaS products: 42% ask about you. The recurring questions, cards vs dropdowns, AI in how-did-you-hear, what to steal.
Of the 304 onboarding screens we classified from 95 SaaS products, 128 are questions about you. That is more than workspace setup, teammate invites, product tours and plan pickers put together, and it makes “tell us about yourself” the defining move in these onboarding UX examples. Invites, the step growth teams talk about most, appear on just 10 screens.
Every screen is on the onboarding screens wall, captured in October 2026 from products such as Lovable, Zapier, Jira and Postman. Below: which steps show up, the questions that keep recurring, how they are asked, and the one product that pays you to answer.
The step mix: mostly questions
We tagged each screen with one step type. The wall has 327 screens; 304 carry a step tag, and that is the base for the counts here.
- Questions about you: 128 screens (42%), from 67 of the 95 products.
- Workspace setup: 50 screens from 34 products.
- Profile details 26, connect or import 25, welcome 20, tour 19, template 14.
- Choose a plan: 12 screens from 11 products.
- Invite teammates: 10 screens from 10 products.
Some invites hide inside other steps. Mermaid Chart puts an email box under two questions with “Or skip - you can always invite later.” Lucidchart slots “Invite your co-workers!” between “About you” and “About your work”. Even so, products spend far more screens learning about the user than recruiting the user’s team.
Onboarding UX examples of the questions everyone asks
We read every question screen and sorted what it asks. Many screens ask more than one thing, so a product can appear in several rows. Counts are products, out of the 67 that ask anything.
| Question | Products | Typical format | Examples |
|---|---|---|---|
| What will you use it for (goal or use case) | 32 | Icon tiles, often multi-select | ElevenLabs (10 tiles), Typeform, Confluence |
| Role or kind of work | 26 | Chips or a card grid of 8 to 14 options | Lovable (8), Zapier (10), Jira (14) |
| Team or company size | 18 | 4 to 6 ranges, or “Just me” vs “With my team” | Zapier, Postman, Any.do, Craft |
| Work, personal or school | 14 | Three cards | Deepnote, Descript, Mural |
| How did you hear about us | 13 | Chip cloud of 5 to 16 sources | Hygraph (16), Runway (12), Make (9) |
| Experience level | 7 | Labelled ladder | Retool, Krea, Mural |
| Tools you use today | 5 | Logo chips | Canny, Penpot |
| Industry | 4 | Chips or dropdown | Smartsheet, SurveyMonkey |
The Work / Personal / Education split is the closest thing to a standard. Deepnote, Descript and Mural use exactly those three words, SurveyMonkey and Penpot use them in a different order, and Gamma writes “For personal use / For work / For education”. Of the 14 products that ask it, 13 offer a school or education answer; Sanity swaps it for “Agency” and Netlify adds “Client”.

Size questions agree on the idea and nothing else. Zapier’s smallest bucket is “1-49”; Postman’s is “1 member”; Gamma runs from “Just me” to “5001+”. Four products (Any.do, Craft, Todoist, Krea) skip ranges and ask only whether you are alone or with a team. A few give a polite exit: Rive lists “I’d rather not say” as a role, and Glasp, which also asks for your highest degree, offers “Prefer not to say”.
“How did you hear about us” now has an AI answer
Thirteen products ask how you found them, and nine show their answer list on screen (the rest use a closed dropdown or free text). Eight of those nine include an AI chat or AI search option. The ninth, Typebot, does not.
Placement varies. Make lists “AI” first, above Search engine, and VEED also opens with “AI chat”. Bolt puts “AI Search (ChatGPT, Perplexity, Claude)” second, right under “Google / Search”. Elsewhere it sits mid-list: fourth of 11 at Bubble, fourth of 16 chips at Hygraph, fourth of five at Evernote (“ChatGPT or another AI tool”), eighth of 12 at Runway, between Reddit and Instagram. Hygraph’s optional text box even uses an AI path as its placeholder: “e.g. Asked ChatGPT for a headless CMS, then a colleague recommended you”.


Cards beat dropdowns, and many drop the Next button
108 of the 128 question screens (84%) answer with cards, chips or tiles. Dropdowns survive mainly where a product crams several questions onto one form: Relume asks five things with four dropdowns, Prezi uses “Occupation” and “Department” selects, and Kapwing tucks “What kind of work do you do?” into a menu.
The bigger shift is what happens after the click. In at least 15 products a question screen has no Next or Continue button at all; the answer card is the button. That includes Lovable, Make, Mural, Gamma (three questions in a row), Jira’s 14-row “What kind of work do you do?”, Zapier’s role and company-size chips, three SurveyMonkey steps and four Runway steps.

The pattern has a clear rule. Single-choice questions drop the button. Multi-select ones keep it: Jira’s “How does your team plan to use Jira?” checkboxes, ElevenLabs’ “Select all that apply” tiles, Plane’s “Select one or more” and Maze’s intent list all end in Continue. Deepnote and Krea are the exceptions that keep a button on single-choice cards.
Three questions, one sentence
Three products turn the form into a sentence. Clockify: “I’m here to [Select primary goal] with my team of [Select size] and our primary billing model is [Select model]”. Postman: “I’d like to [Select option] as a [Select role]”. Smartsheet: “My company has [1 – 49] people / I am in [Technology]”.

Smartsheet goes further. Picking a goal chip writes an editable AI prompt (“Create a simple Project Timeline for the Product Development team…”) beside a workspace preview labelled “Preview includes data generated by AI.” The answers stop being survey data and become the first piece of the product.

Progress shown more often than skip
Across all 304 screens, 150 show progress (49%) and 98 offer a skip (32%). On question screens the gap widens: 66 of 128 show progress, only 35 let you skip. Sixteen of the 95 products show neither on any screen we captured.
Progress comes in many dialects: “Step 2 of 3” (Superlist), “3 of 5” (SurveyMonkey), “Question 2 of 5” (Glasp), “2/6” (Krea), dots and bars everywhere else. Skip copy is more inventive. Grafana Cloud writes “Know what you’re doing? Skip this and show all options”; folk declines invites with “No, I prefer to work solo”; Grammarly just says “Later”. Mural’s questions offer Skip but no Next at all.
Bolt pays 1M tokens for three answers
Bolt is the only product in the set that puts a price on the questions. Its survey arrives as a modal over the builder after the first app is generated, with “Get 1M tokens for completing the survey!” in the header of all three steps (1/3 to 3/3). The last screen reads “1M extra tokens have been added to your account!” and offers another: “Want an extra 1M tokens? Drop your work email”.
Two things are worth noticing. Bolt asks after you have seen value rather than before, and it pays in its own currency, a reward that costs Bolt compute rather than cash. See the reward screen and the full Bolt collection.
What to steal
- Ask what changes the product. Use case, role and size cover most of the set. If an answer does not change the next screen, cut it.
- No Next button on single-choice questions. Let the card submit, as Lovable and Zapier do. Keep Continue for multi-select.
- Add AI to your source list. Eight of nine visible lists include it. Copy Bolt’s optional “Which AI tool?” follow-up to learn which assistant sends you users.
- Collapse short questions into a sentence. Clockify fits three questions on one line.
- State the cost up front. Postman’s “1 min” and “2 steps” does more than a progress bar.
- Use the answers immediately. Smartsheet turns them into an AI prompt and a preview; Bolt asks only once you have built something.
How we counted
We captured these screens in desktop browsers with research accounts in September and October 2026. A vision model recorded each screen’s step type, layout, progress, skip and card attributes, and you can check each one on the screen’s page; we read every question screen ourselves to sort the questions in the table, so those counts are our classification. We record what we saw on one pass through each flow, not A/B tests, so nothing here says which pattern converts better. Flows can differ by account, region and plan.
To build your own flow from these references, browse the onboarding screens wall or connect the Screenrove MCP server and ask your agent for “onboarding question screens with no Next button”.