GIMMEFY LABS · DAILY TEST
PARTIAL
28 JUL 2026 · LAYER 03 FAILED
the daily run · issue Nº 003 · 28 july 2026

SnapStudy.ai refuses to give your child the answer. Six AI models think it does anyway.

Four audits and six expert personas worked one question for a morning. Search told this brand to become an answer engine. The machines said it already was one. The humans said it was the best-loved thing in its category. All three were looking at the same product.

One brand. One morning. Every part of our AI infrastructure, tested in public, including the part that broke.

A girl working a maths problem at a kitchen table at 9pm while a parent walks away, under the headline Clues, not answers
Made in Visual Hub with the SnapStudy.ai brand vault switched on. The on-image line was specified exactly and rendered first time. Open it →

The 60-second version

  1. The question. Can an AI homework app win by refusing to give the answer?
  2. The answer. Yes, but not as currently built. Six expert personas deadlocked three against three, then designed a way out that neither side had proposed: keep withholding the answer, but deliver the clue as a 15-second yes-or-no tap instead of a paragraph.
  3. The surprise. Its single biggest organic search opportunity is "200 celsius to fahrenheit", 201,000 searches a month. A pure lookup. The exact thing the product refuses to do.
  4. The problem. Six frontier models were asked to recommend a homework tool. None of them named SnapStudy. Four of its five positioning claims came back marked contradicted, and the machines handed its own differentiator to a competitor.
  5. The cost. 555 credits and one morning. One capability failed outright and is written up in section 7.

Every link below opens the real run. No account needed. The 12-page PDF is here →

Nº 003

Take the whole issue with you

Twelve pages, designed to be read on a train or forwarded to a colleague who will not click through eight sections. Same story, same links, none of the scrolling.

Download the PDF →

Section 1

What this is about.

1.1  What we are doing

We build AI infrastructure for marketing teams. Four layers. Every morning we test all four, end to end, to be sure nothing we shipped yesterday broke anything.

Layer 01Models  48 frontier models, routed for you
Layer 02Context  brand vaults, pulses, the memory that makes output sound like you
Layer 03Capabilities  the workflows, media and research that do the work
Layer 04Governance  workspaces, roles, audit trails, white label

A test needs a subject. So each day we pick a brand we admire and run it through everything. Today: SnapStudy.ai, a camera-first homework app for 10 to 14 year olds, and a question its team could genuinely be sitting on this week.

Today's question

Can an AI homework app win by refusing to give the answer?

It is a real question because the opposite bet just collapsed in public. Chegg sold answers and lost 31 per cent of its subscribers, 30 per cent of its revenue and roughly 99 per cent of its share price to free chatbots. SnapStudy does the reverse. In its own words the AI "provides clues for you so that you can uncover each step of the problem-solving process on your own", and parents and educators set "the level of assistance that child receives". Clues, not answers. Adults hold the dial.

There are three ways to drive the platform. Each issue runs through one, so all three get tested in public. Here is what each actually looks like.

Today's path · tested in this issue
The gimmefy platform: every capability down the left, brand context along the top
1. The platform

Every capability down the left, your brand context along the top. Pick a tool, brief it, watch it work. Everything in this issue ran here.

Friday asking what it can get done for you, with a one-line brief box
2. Friday

Say what you want in a sentence. Friday reads it, proposes a plan with costs, waits for "Go", then does the work. Scheduled for today and pulled, for the reason in section 7.

Claude, ChatGPT and Gemini connecting through the gimmefy connector to your brand vault and every capability
3. MCP

Connect gimmefy inside the AI client you already use. Your vault and every capability, without leaving the chat. Not available on this instance yet, and section 7 says so plainly.

How this got doneWe built a brand vault for SnapStudy.ai from public material only, then sent four environment reports out to study the brand's world. Everything after that reads from those two things. No prompt in this issue starts from a blank page.

1.2  What is in this report

It runs in the order the morning ran. What the machine learned before we asked it anything, what we asked, what came out, the single best piece of work, then the rest of the day capability by capability. Section 7 is how it was built and what broke. Section 8 is every link.

SnapStudy.ai is not a gimmefy customer and has not seen this. We picked them because the question they face is a genuinely hard one and we wanted to show what the platform does with it. Everything here was generated on gimmefy from public sources and is unofficial. Campaign concepts, forecasts and financial figures are model output, not the brand's real activity or measured results.


Section 2

Set-up and discovery.

Before asking the machine to make anything, we ask it to go and learn.

It starts with a brand vault. Think of it as a 360-degree understanding of a business held in one place: who the brand is, how it sounds, what it sells, who it sells to, what it looks like. Not a folder of files. A context layer every capability reads from, so nothing downstream has to be told the brand twice. Ours came back with 38 of 38 fields filled and marked grounding-ready, in under two minutes.

Then four environment reports go out and study the brand's world, and their findings fold back into that same vault. Listening reads the public conversation. Search reads what the website earns. GenAI asks the chatbots what they say about the brand. Landscape scores the brand's claims against what the market actually sees.

2.1  Listening

113 pieces of public evidence read across six kinds of source, share of voice measured against every rival, sentiment scored for each.

BrandShare of voiceNet sentiment
Khanmigo30.1%+0.41
Chegg29.2%+0.21
Photomath22.1%+0.32
Socratic by Google9.7%+0.36
SnapStudy.ai8.8%+0.67

The finding

Last in awareness. First in affection, by a distance. The people who have actually found this product like it more than anyone likes the market leaders.

Nobody has heard of it. Everybody who has, likes it.

A limit worth stating. 112 of the 113 items carry no date. Treat the twelve-month window as the collection window, not a trend line. Nothing here supports a claim about momentum in either direction.

2.2  Search

A live crawl plus keyword, backlink and Lighthouse pulls, scored across five dimensions.

Content & relevance
100
Technical health
95
On-page SEO
80
Performance
69
Authority
33

Findability lands at 74 out of 100. Content scores a perfect 100 on 1,183 ranking keywords. Authority scores 33, because the whole site rests on 19 referring domains. And exactly one keyword sits in the top three positions: the brand's own name.

The finding that starts the argument

The two biggest content gaps this audit found are "how many weeks in the year" at 823,000 searches a month and "200 celsius to fahrenheit" at 201,000. Pure lookups. No method to teach, no step to uncover. The search channel is telling a company built on refusing to answer that its largest growth opportunity is to answer.

Thirteen of its fifteen near-miss keywords are definition lookups. The pattern is not an accident.

2.3  GenAI

Six frontier models asked the same buyer questions independently, the first one deliberately never naming the brand.

FGrade, 31 out of 100
0 of 6Models that named it unprompted
0%Share of the AI conversation
4 of 5Positioning claims marked contradicted
DimensionWhat the brand saysWhat the AI says
PositioningA guided companion giving step-by-step cluesA "snap-and-solve answer engine that provides full solutions immediately"
Proof pointsExplicit parent and educator controlsControls are "essentially nonexistent"
Competitive framingThe ethical alternative to answer-vendingHands that exact differentiator to Khanmigo

The sentence the whole issue turns on

The company's entire reason for existing is the one thing the machines do not believe about it.

Read the grade with this attached. Three of the six engines, Claude, Perplexity and Kimi, returned no usable data. The report says so itself. The F effectively rests on three engines, and we are not going to print it bare.

2.4  Landscape

The category, the rivals, and how far the brand's own account of itself sits from the market's.

B−Overall standing, 65 of 100
85Positioning clarity
30Efficacy and proof
35Category attractiveness

Positioning clarity is the highest score on the board. Proof is the lowest. The report's own summary of the gap is blunt: the brand "asks parents to pay for a scaffolding tool over a free answer engine, yet reveals zero verifiable proof that students actually learn more."

It also named the vacant ground: SnapStudy sits "in the highly lucrative, vacant middle between Photomath's high-speed cheating risk and Khanmigo's high-friction institutional trust."

Clearest story in the category. Thinnest evidence for it.

Why this exists. You can read your own reviews. You cannot read what a chatbot says about you in a stranger's private window, and you cannot see the queries you are invisible for. These four reports are how a brand finds out what the market and the machines already think, before it spends a penny talking to either.

Section 3

What we asked it to do.

Picking is the skill. There are dozens of things the platform can run, and the useful move is choosing the four or five that answer the actual question.

The question was a decision, not a task, so the choices followed from that. Understand the market, use Beacon. Pressure-test the decision with people who disagree, use Focus Group. Turn a decision into work, use a playbook. Make it real, use media. Each brief below had to cite a fact from the reports above, and each carried one hard constraint so we could check afterwards whether the machine actually obeyed it.

The debate

Six expert personas, one decision: hold the line on clues, or drop it. Constraint: no fence-sitting, no "do both", everyone must say "clues, not answers", and they must attack each other by name.

Read the debate →
The campaign

A US launch aimed at parents who quietly worry AI is doing their child's homework. Constraint: never use the words cheat or cheating, and make the parent's assistance dial a hero, not a footnote.

What happened to it →
The key visual

The 9pm kitchen table, the moment before understanding. Constraint: exact on-image text, rendered once, in the brand's own deep teal.

See the visual →
How this got doneEvery brief quoted the reports back at the machine. The debate brief carried the F grade, the 201,000-a-month lookup, and the +0.67 sentiment. That is why the arguments below are about this brand and not about homework apps in general.

Section 4

What we got done.

4Environment reports, six minutes
6Expert personas, one decision
555Credits, whole morning
1Capability that failed
  • The recommendation. Hold the line on "clues, not answers", and change how the clue is delivered. Not a strategy pivot, a UX one.
  • The operational half. The F grade is a metadata problem, not a business-model problem. Publish a governance page and an llms.txt that state the pedagogy in machine-readable terms, and the misclassification has something to correct itself against.
  • Why it matters commercially. The buyer and the user are different people. The parent pays for the friction. The child experiences it. Every decision in this category is really about that split.
  • The unexpected finding. The search channel wants this brand to become the thing it exists not to be, and it is worth 1,024,000 searches a month to say no.
  • Assets produced. One brand vault, four reports, one debate with synthesis and decision matrix, one key visual, two public share links.

Section 5

The star today: six experts who could not agree.

The best thing the platform produced today was not an asset. It was an argument that refused to resolve, and then resolved itself in a way nobody in the room had proposed.

Why this exists. Almost all AI use is one human talking to one model. That is fine for a task and risky for a decision, because you cannot tell a confident answer from a correct one. A focus group is one human to many, and the many are briefed to disagree.
The Focus Group: the decision on the table, six expert cards each carrying one line of their argument, and a footer explaining the facilitator and synthesiser
How a focus group is set up: the decision at the top, the experts you chose, and a synthesiser that has to reconcile them at the end.

We picked six lenses that map to the people who really decide this: the CEO, a CFO, a cultural strategist, a devil's advocate, a consumer psychologist and a UX strategist. Five rounds, observer mode. The brief agent sharpened our question before it went to the room, and its version was better than ours:

Does holding the pedagogical high ground make you a category-defining mentor, or just an unnecessary friction point between a 12-year-old and their finished homework?The brief agent, rewriting our question

5.1  They split three against three

The fault line is the interesting part. The three who think about the buyer said hold. The three who think about the user said drop.

Aris and Marcus are designing a product for a user who doesn't have a credit card.The CFO, holding the line
At 9:00 PM, a 7th grader's prefrontal cortex is exhausted. We aren't scaffolding, we're creating an artificial wall that guarantees app abandonment.The consumer psychologist, against
The UI feels like a fast digital camera tool, but the UX acts like a strict school teacher.The UX strategist, against
The snap-to-solve market is toxic radioactive waste.The cultural strategist, holding the line

5.2  Then it designed a third option

The synthesis rejected both opening positions and approved something neither camp had put on the table.

OptionVerdict
Pivot to answer engineRejected. "The exact path that vaporized Chegg's market cap."
Hold the line, current UXRejected. "The positioning is correct, but the execution is fatal."
Hold the line, progressive disclosure UXApproved. "The only path that satisfies both the exhausted student and the paying parent."

Reveal number one

Keep refusing the answer. Change the shape of the clue. Replace a paragraph of Socratic prose with a 15-second binary tap, "did you simplify this fraction first?", so the tired child gets momentum without ever being handed the destination.

It also produced the most useful reframe of the morning, which costs nothing to act on: the F grade is "a content indexing bug, not a business model flaw." Fix the schema, not the strategy.

Watch all six argue it out →

5.3  And then it attacked our own headline

We led this issue on "last in awareness, first in affection". The devil's advocate turned that straight back on us.

That +0.67 net sentiment might be a textbook case of survivorship bias. Because it has the lowest share of voice in the market, that high score may only reflect a tiny, hyper-compliant fringe of parents, while ignoring the silent 91.2% majority of students who hit the friction point, quietly churn, and open a free chatbot.The devil's advocate, on our cover story

Reveal number two

Our two best findings might be one finding wearing two hats. We have left both in, next to each other, because that is the honest position and because the machine caught it before we did.

How this got doneSix personas chosen from the roster, five discussion rounds, observer mode, one mid-debate instruction forcing them to cross-examine the F grade against the sentiment score. The synthesis, decision matrix and open questions are generated at the end from the transcript, not written by us.

Section 6

Everything else that ran.

6.1  The brand vault

What we asked. Build a 360 of SnapStudy.ai from its public site and nothing else.

What came back. 38 of 38 fields, 98 per cent complete, marked grounding-ready, in under two minutes, with 13 visual assets found on the site. It independently landed on the framing that matters: "unlike many homework-help brands framed as answer engines, SnapStudy.ai consistently markets itself as a guided learning companion."

Why it is useful. Everything downstream reads from it. The debate above argues about this brand specifically because the vault told it who this brand is.

Two honest notes. The vault captured zero logos and still reported its visual health as healthy, which is generous. And we deliberately removed two headshots of real people from the asset set before anything generated against them.

6.2  The key visual

What we asked. The 9pm kitchen table, the moment before understanding, parent deliberately walking away, with one exact line of on-image text.

What came back. The image at the top of this page, first attempt, five credits, with the line rendered correctly in the brand's own deep teal.

Why it is useful. On-image text is where most image models fall over. Specifying it exactly, once, in a named colour, is the difference between an asset and a redraw.

Open the visual →

6.3  The campaign that did not finish

What we asked. A full US launch campaign, 21 stages, with two banned words and a mandated hero mechanic.

What came back. Seven stages of genuinely good strategy, then a wall. The brief, three intelligence stages and three strategy stages completed. It named the campaign "Clues, Not Answers" and wrote the line the whole thing hangs on:

We aren't selling an app that does less. We are selling the end of the parent's role as the homework referee.Campaign Creator, stage 1

Our hard constraint survived all seven stages word for word, which is the thing we were actually testing: "never use the words cheat or cheating", "the assistance dial must be shown and named as a hero mechanic", "zero claims about measured learning outcomes".

Why it is useful anyway. A playbook that fails at stage 10 still leaves you seven stages of work. Section 7 has what broke and why.


Section 7

How it was built, and what broke.

555Credits spent
6 minFor all four reports
7 of 21Campaign stages completed
3Bugs filed or updated

Everything ran on the platform surface, inside one project, on a white-labelled instance of our own. The vault was built from the public website only. That last point matters more than it sounds: this subject reached our attention through a partner's account, and we deliberately did not read a single thing from it. It is the only reason the line "not a customer" at the top of this page is true.

7.1  The gap we found and logged

  • The campaign died and stayed dead. Every creative stage needs a key visual, none could be produced, and three resume attempts failed identically. Logged against an existing critical ticket. The same image model works perfectly on its own, which is how we know it is the playbook's media path and not image generation.
  • Two browser tabs held two different projects at once. Same account, same session. Work submitted from the stale tab would have filed itself into yesterday's project with no warning. Caught before the debate ran.
  • The bulk report refresh ignored its first click, then did not repaint when the reports finished. We only knew they were done because we went looking in the database.

7.2  The thing that made us wince

The debate synthesis is beautifully argued and its financial spine is entirely invented. None of these came from us, from the vault, or from any source, because SnapStudy's pricing, costs and churn are all private:

  • $270 lifetime value on a $15 a month subscription
  • Parent churn rising 3% to 12%
  • GPU cost falling $0.08 to $0.03 a session, expanding margin to 84%

The auto-generated infographic brief then promoted two of those inventions to headline statistics. This is not an argument against the tool. It is a precise description of where the human belongs: the reasoning holds up, and every number needs checking before it leaves the building.

The lesson we are keeping

Print the strategy. Strike the arithmetic. Say why.

7.3  One thing that made us laugh

We filed a bug accusing the platform of running three of four reports. It had run all four. The fourth simply uses a different executor and was invisible where we went looking. We rewrote the ticket with the retraction attached and downgraded it, which is roughly what we would want anyone else to do.

7.4  What we did not run

Model Council, video and staged publishing did not run today. The Daily Run's own instance has no MCP endpoint yet, so nothing in this issue was made through the connector and we are not going to pretend otherwise. Today's stamp says PARTIAL for that reason.


Section 8

Appendix: everything, with links.

Listening reportLast of five in share of voice at 8.8%, first of five in net sentiment at +0.67.
Open the report →
Search auditFindability 74 of 100. Authority 33 on 19 referring domains. The answer-query contradiction.
Open the report →
GenAI auditGrade F, 31 of 100, 0% share of AI voice, four of five claims contradicted.
Open the report →
Landscape reportB minus, 65 of 100. Positioning clarity 85, efficacy proof 30.
Open the report →
The debateSix experts, split three against three, synthesis approved a third option neither camp proposed.
Read the debate →
Key visual"Clues, not answers", generated first time with exact on-image text.
See the visual →
The PDFTwelve pages, the whole issue, designed to be forwarded.
Download →
Method note. The market figures the platform cites back to us, Chegg's decline, the 53% and 62% usage figures, the $2.7bn category size, are figures we supplied into the vault from public sources at the start of the morning. The platform reasoned over the evidence it was given rather than discovering it independently, and it should be read that way.
Nº 003

The whole issue, twelve pages

Everything above, laid out to read on a train or forward to someone who will not click through eight sections.

Download the PDF →

Made on gimmefy: Brand Vault, Beacon (Findability, GenAI, Listening, Landscape), Focus Group, Visual Hub. Run on a white-labelled instance of our own, which is the same platform wearing our brand. MCP included: use gimmefy context and capabilities from ChatGPT, Claude, Gemini or any compatible AI client.

SnapStudy.ai is not a gimmefy customer and has not seen this. We picked them because the question they face is a genuinely hard one and we wanted to show what the platform does with it. Everything here was generated on gimmefy from public sources and is unofficial.