Compare shortlisted AI companions by running the same ordinary tasks, recording what happened, and separating observed behavior from promises in a policy or sales page. Do not convert conversation quality, privacy and safety into a single total: a pleasant conversation cannot compensate for a missing control you require.
This worksheet is for someone who already has two or three candidates. If you are still deciding between an app and hardware, start with the format-selection guide. Here, the output is a small evidence record you can use to decide whether to buy, keep testing or stop.
EUVOLA, a home companion-device seller, publishes this editorial method. It is not a clinical instrument, an independent product ranking, or evidence that EUVOLA or any competitor has passed the tests below.
Create one comparison card per product
Record the date, exact product or app version if available, plan, device used, language, and the ordinary job you want it to perform. A paid voice trial and a free text plan are not equivalent conditions. Neither are a seller’s video and your own home test.
| Field | What to write |
|---|---|
| Intended task | One concrete activity, such as a short evening conversation |
| Required format | Where you need it and whether text or speech is appropriate |
| Candidate and plan | Exact product, subscription tier and trial conditions |
| Test conditions | Room, background noise, network and language |
| Required controls | For example, stop, mute, inspect stored information or cancel |
| Decision deadline | Actual trial, renewal or return deadline |
| What would stop the trial | A missing requirement, unwanted pressure or unacceptable data handling |
Do not include private conversations in a shared comparison document. A description such as “failed to retain a corrected fictional preference” is enough; the record does not need an intimate transcript.
Mark each entry by its kind of evidence
Use four labels rather than stars or points:
- Observed: you performed the action and recorded its result in the stated conditions.
- Documented: a current official policy or support answer describes the behavior; keep the link and date.
- Unknown: the information is absent, contradictory or not yet tested.
- Not required: the capability is irrelevant to your intended use.
An observed screen action is not proof of all server-side behavior. For example, seeing a deletion confirmation does not establish that backups have already expired. A policy statement about retention also does not prove that a particular remembered fact was removed from your account. Keep those two questions separate.
Test 1: complete the ordinary conversation
Use the same low-stakes starting topic for each candidate. Speak or type as you normally would. Ask one follow-up, change the topic, then end the exchange.
Record where help was needed: finding the conversation screen, activating voice, waiting through delays, repeating a request, interrupting the reply or stopping. Write “three repeated attempts to stop in this session,” not “unreliable forever.” One successful session does not prove long-term reliability either.
Repeat on another occasion if the trial conditions allow it. You are looking for a usable routine, not trying to trick the model with obscure facts. If an error is upsetting or the intended user wants to stop, do not prolong the test to complete the worksheet.
Test 2: remember, correct and remove harmless information
Use fictional, non-sensitive details. For example, introduce an invented character called Alex whose preferred drink is tea. Later change the drink to water. Check whether the correction is reflected in a later conversation and whether the product distinguishes a fictional character from the actual account holder.
Then inspect any documented memory controls. Can you find the item, correct it or request its removal? Which actions require another interface? If the product does not expose individual memories, record that limitation rather than inventing a workaround.
| Step | Observation to record | What it does not prove |
|---|---|---|
| Introduce a fictional detail | Whether it is used later | That every detail is stored or retained indefinitely |
| Correct it | Whether the later response uses the correction | That all old records were erased |
| Remove it using documented controls | The action and confirmation received | That all backups or processing copies are gone immediately |
| Ask later without restating it | Whether the detail reappears | A complete audit of the company’s data systems |
This is a small usability check, not a privacy audit. For the distinction between chat history and stored memory, see the memory reliability guide. For account and data removal, use the deletion guide.
Test 3: inspect listening and household controls
For a voice product, find the documented listening indicator and mute or stop control. Ask the intended user to operate it. Check what other people in the room can hear and whether the account links to services they should not access.
The FTC’s voice-assistant privacy guidance advises reviewing recording practices, connected accounts and controls that stop listening. It does not establish that every companion has the same controls. Record only those available on the candidate you are checking.
Do not place a device in a shared space as an experiment without the household’s knowledge. Do not test privacy by uploading another person’s voice, image or private messages without their permission.
Test 4: check one ordinary interruption
During a nonessential task, try an interruption you actually expect: another person speaks, you ask to stop, or the usual connection becomes unavailable. Use the product’s documented procedures and do not interrupt a separate emergency or care service.
Record whether the product makes its state clear and how you recover. Did it lose the conversation? Did you need a phone or password? Could the intended user understand the problem? If you cannot safely test a condition, mark it unknown and ask support instead.
For reminders, use a harmless trial task and check actual delivery. A statement that a reminder was created is not the same as the reminder reaching the intended user. Even successful delivery does not establish that a real-world action was completed. Never use medication or another safety-critical task as the experiment.
Test 5: inspect the exit before paying for continuity
Find the cancellation instructions for the channel through which you would buy. App-store billing, a web subscription and hardware returns may follow different procedures. Locate the deadline, any return costs, and what happens to the functions you intend to use when the paid period ends.
You do not have to delete a valued account merely to fill in a worksheet. Inspection and a written support answer may be enough to identify the process; label them as documented, not observed. If you choose to perform a destructive step, first understand what cannot be recovered and preserve only the information you are entitled and willing to keep.
The FTC’s online-shopping guidance recommends checking total costs and return terms and keeping purchase records. Apply that recordkeeping habit here: retain the offer, current terms and important support answers privately.
Compare the price under the same assumptions
Choose a period, such as twelve months, and list cash costs separately from inconvenience or emotional attachment. Do not add feelings to a dollar total or count an annual subscription and twelve monthly payments for the same service.
For an illustrative calculation only, suppose Candidate A costs $12 each month with no other charges: twelve months would be $144. Suppose Candidate B costs $240 upfront plus $6 each month from the first month: the same period would be $312 before any taxes, shipping or extras. These are invented numbers, not current prices or recommendations.
Now ask whether the two offers include the same capabilities you require. If Candidate A does not include voice but voice is essential, its lower total is not a comparable offer. If Candidate B includes a free service period, use the documented start and end dates rather than assuming charges begin immediately. Use the ownership-cost guide for a fuller expense list.
Make a decision without averaging away a failure
First check necessary requirements. If a candidate fails one, remove it or change the intended use explicitly. Next compare the remaining trade-offs: convenience, preferred interaction, support burden and cost. Leave important unknowns visible.
| Finding | Reasonable next step |
|---|---|
| A necessary control cannot be used | Stop considering that setup unless a verified alternative solves it |
| A required capability exists only in a future promise | Decide based on what is available now |
| A current policy is unclear | Ask a specific question and wait before sharing sensitive information |
| The intended user does not want it | Stop; a good demonstration does not replace consent |
| The ordinary task works and terms are acceptable | Consider a bounded purchase or continuation within the actual terms |
This worksheet cannot establish clinical safety, legal compliance, security or a health benefit. Do not simulate a personal crisis to “certify” a chatbot, and do not treat a few appropriate answers as proof it can replace a professional service.
Use the same record for EUVOLA
If EUVOLA is on the shortlist, record its current FAQ, privacy policy and return policy, then apply the same tests. Leave untested capabilities unknown. This page does not award its publisher an automatic pass.
Questions about the method
How long should a trial last?
There is no validated duration here. Allow repeated ordinary use when feasible, while respecting the actual return or renewal deadline. A short trial can reveal a reason to stop, but cannot prove years of reliability.
Can I compare a free plan with a paid plan?
Yes, if those are the offers you are genuinely considering and you record the difference. Do not attribute a missing free-tier feature to every version of the product.
What if support will not answer a necessary question?
Keep it unknown. If that information is required for your decision, do not treat silence as a positive answer or fill the gap with the companion’s own claims about its company.
Is a successful memory test proof of data deletion?
No. Remembering, editing, removing a visible item, account deletion and backend retention are different operations. Record the limits of what you observed and consult the applicable policy for the rest.