What the extension sends, and what it never touches
This policy covers the Cresva Chrome extension only. It supplements our general privacy policy, which governs cresva.ai and the signed-in product, and adds to it rather than replacing it. Where the two speak to the same thing, the stricter statement is the one we hold ourselves to.
Last updated November 16, 2026
What a scan sends
Nothing is sent until you click. The first time on a new install, the panel waits for you to press Inspect this store; after that, clicking the Cresva icon on a page inspects it straight away, and you can switch that off in the panel. There is no background collection, no scan on page load and no scan of tabs you are not looking at.
Always, when you scan:
- The address of the tab you scanned, stripped. The query string is reduced to
variantandsku, and the fragment is dropped, in your browser, before the request is assembled. Session tokens, affiliate ids, click ids and email-campaign parameters all live in query strings and none of them reaches us.variantandskustay because they select which product the page is showing, so dropping them would point us at a different item than the one you are looking at. - A page kind, which the extension works out from the address alone: one of
product,collection,homeorother. - A random install identifier and a proof of work value, sent as request headers. Neither identifies you. See Anonymous by default.
Additionally, when Read this page in my browser (the browser view option, under What Cresva reads in the panel) is on. It is on by default for a new install, and nothing is sent until you press Inspect this store under a line saying the page is read in your browser. An install updated from a version before 5.0, where it was off, keeps it off until you choose. Whatever you choose stays. With it on, an inspection sends:
An extract of the page, read from what your browser has already rendered. This is what lets the panel tell you that something a shopper can see is missing from what an agent is served. It contains:
- the page title and its H1 headings
- structured data: the bodies of the page's JSON-LD blocks, up to 32 blocks and 32KB in total, plus the
@typeanditemtypenames found on the page - meta tag names and values, up to 40 of them
- price shaped strings visible on the page, availability text, and whether an add to cart control is present
- counts only: how many words are visible, how many script tags there are, how many JSON-LD blocks there are, and how many bytes the extract came to
- up to 2,048 characters of visible text from the page's main content region
- the page address a second time, this one reduced to a whitelist of
variantandsku, with everything else in the query string and the fragment dropped - the country the page says it is priced for, when the store's page states it
- on a Shopify store, the store's public product list, read from your browser the way the page reads it: for each product its name, link, publish date, type and lowest and highest price, with the currency the page shows. At most 48KB. No descriptions, images, cookies or anything about you. When the page shows you are signed in to the store, the list is not read from your browser: Cresva reads it from its own servers instead, so no customer, wholesale or member prices reach Cresva
Before any of that leaves the page, three redaction passes run in your browser, not on our servers. Email addresses become [email removed]. Digit runs that are the length of a payment card and pass the Luhn check become [number removed]. Phone shaped strings become [phone removed]. These passes run on everything listed above: the title, the headings, the text sample, every value in the structured data and the meta tags. Two kinds of value are spared the passes that would break them: product identifiers such as barcodes, SKUs and part numbers are kept as published, and web addresses get the email pass only, because a long product id inside a link can look like a card number. In the free text sample only, any remaining run of four or more digits is also removed, with prices deliberately kept because prices are the thing the scan exists to measure.
The text is read from what your browser rendered. The only markup read is the structured data blocks and the named meta tags above: no other script bodies, no inline event handlers and no hidden attributes travel with it.
What is never read
- Cookies. The extension does not request the
cookiespermission, so Chrome does not give it the ability to read them, on any site. Every request it makes to Cresva's servers is sent with no cookie attached, so no cookie value reaches Cresva; the only sign-in it carries is the token Cresva itself issued, once you sign in. - Local storage, session storage and IndexedDB. The extension reads none of them on any page you visit. The
storagepermission it does hold writes only to the extension's own storage, and only the entries listed in full under Anonymous by default below. Three of them are ever sent to us: the install identifier, and, once you sign in, your Cresva sign-in and the client identifier it was issued to. - Credentials. No passwords, no session tokens, no authorization headers and no API keys are read from the pages you visit. Signing in to Cresva is a separate, optional step for the ads panel and for analysing Ad Library ads, described under Signing in below. Nothing a page carries is ever read as a credential.
- Form fields. Before any text is read, inputs, textareas, selects, options, buttons, forms and contenteditable regions are removed from a working copy of the page. Whatever you have typed lives in exactly those elements, and they are gone before the text extraction starts.
- Other tabs. The extension does not hold the
tabspermission. Access is granted one tab at a time by your click on the Cresva icon, and it does not carry over to the next tab or the next page. - Anything at all, until you click. The extension declares one host permission,
https://cresva.ai/*, which is our own API. It holds no standing permission on any store you visit. Its one standing script runs only on the Meta Ad Library, where it places an Analyse with Cresva button on each ad and reads nothing until you press one; see Analysing an ad below.
Pages the extension refuses outright
On these pages no request leaves your browser at all. Not a redacted request, and not a scan with the browser view turned off. Nothing. The refusal happens before a request is even assembled, because the address by itself is a disclosure: a Shopify admin URL tells us which store somebody administers and which order they were reading.
Refused by host:
admin.shopify.com, partners.shopify.com, accounts.google.com, myaccount.google.com, admin.google.com, login.microsoftonline.com, adsmanager.facebook.com, business.facebook.com, ads.google.com, ads.tiktok.com, analytics.google.com, app.klaviyo.com, mail.google.com, and any subdomain of those.
Refused by path, anywhere in the address:
admin, wp-admin, wp-login, administrator, checkout, cart, account, orders, order-status, my-account, profile, addresses, payment and billing. File extensions are stripped before matching, so /wp-login.php and /admin.aspx are refused too.
Refused because of what is on the page:
A store locked behind its password page. The address /password is refused before anything is read. A page at any other address is refused when it has a password field and is the store's password gate: the page carries Shopify's template-password marker, or a password form that posts to the store's own /password. That cannot be decided from the address, so it is checked after the page has been looked at, from the page's markup only; no field's value is ever read. When it fires, the extract already held in the extension's memory is discarded unsent and unlogged, and the scan ends there. A page that merely contains a customer sign-in field, such as a sign-in drawer in a store's header, is not refused for that and is inspected as normal; a store's own sign-in page is still refused by its account path above.
Browser pages are also refused: chrome://, about:, devtools://, view-source:, extension pages, and anything that is not http or https.
Signed in, on your own ad accounts, this changes and it changes narrowly:
Everything above describes the extension as anyone can install it, with no account. Cresva reads storefronts, not the screens behind your login. The thirteen hosts listed above are refused outright in that mode and nothing you do makes a scan of them happen.
Once you have signed in and connected an ad account you own, the panel opens on that account's own dashboard and shows you what Cresva already knows about the campaign you are looking at. On your own ad accounts, once you've signed in and connected them, it reads only the page address to know which campaign you're looking at. It never reads the page.
It never reads the page, and that is structural rather than a promise. The extension holds no host permission for any ad account domain, so the script injection the store mode relies on cannot reach an ads page at all. There is no setting that turns it on, and no version of this extension in which it reads one. The refusal list above is unchanged: a scan of those hosts is still refused, and so is a try-to-buy.
Store run
When you start a store run, the store's address is sent to cresva.ai and Cresva may publish a public page about that store, which anyone can remove.
The panel runs it only when you press Run it on this store. Inspecting a store does not run it. It sends the store's domain and nothing else: no page content, no cookie and nothing about you.
- What Cresva reads with it. The store's public catalogue, homepage and one product page, the store's live ads in the Meta Ad Library, and the answers ChatGPT, Claude, Perplexity and Gemini each give to the same ten shopping questions about what the store sells. The questions name the store only when nothing else says what it sells.
- How the store's category is found. Up to 60 of the store's public product titles and its product types are sent to Google (Gemini) to work out what it sells. Its category is found by fixed rules first; only when they cannot place it, up to 40 product titles are sent to Anthropic, then OpenAI, then Google if neither answers. When the catalogue cannot be read, the homepage's title and description, the store's live ad headlines, its name and its address are used instead. The brand names found in the answers are sent to Anthropic, then OpenAI, then Google, to check which are brands. Nothing about you is sent.
- Nothing on the store changes. The run stops at payment: nothing is added to a cart, ordered or charged.
- The public page. It shows what the run found, and anyone can remove it from the page itself, with no sign-in. A store whose owner uses Cresva never gets one.
- What is kept. The run's record: what it read from the public store, how long it took, and a salted hash of your IP address for the limits below, never the address itself. After 90 days, by the same nightly job as scans, the hash is deleted, and so is the record, except the store's latest run, which is kept without the hash.
- Limits. At most 8 runs an hour from one network and 12 an hour for one store.
Try to buy, and where it stops
The buying test follows a product the way an outside shopping agent would, made by us, from our servers. It is worth being exact about what that means, and about where it stops.
Every request our servers make to a store, this one and every read, names itself as CresvaBot and the reads are signed. The CresvaBot page says what it asks for, how a store can verify it and how to stop it.
It runs only when you press Watch an AI shopper buy it on a product page. It never runs by itself and never on any other kind of page, and the number of walks on any one store is capped each hour, across everybody using the extension.
Nothing comes from your browser.
Your browser does not take part in this at all. Cresva's own servers make the requests, from our address, with no cookie, no session and nothing carried over from your tab. The store sees us, not you, and the walk would run identically if you were not signed in to anything.
What we ask the store.
- Whether the store publishes the machine-readable routes an agent would use, which is one request for a well-known address and, where it exists, one call asking that endpoint to list what it can do.
- A product, read from the store's own product route, for its variants, prices, currency and stock.
It stops before the cart, and nothing is created on the store.
On a store Cresva does not host, we put nothing in its cart: no cart, no checkout, no emails and no stock held. Every request is a read. A checkout is never started and never completed. No card, no token and no wallet exists anywhere in this path, and the agent profile we publish declares an empty list of payment methods so a merchant can verify that rather than take our word for it.
On an agent storefront Cresva hosts for a merchant, the test goes one step further on Cresva's own servers: it opens a transaction marked as a dry run, asks for shipping and payment options, and never confirms it.
A link to a check you made
When you press Copy link or Send to developer, Cresva publishes the check in front of you at its own address, cresva.ai/check/ and a random code of 24 characters. Anyone you give the link to can open it; nobody can find it without the link, because the code cannot be guessed and the page asks search engines not to index it.
The page holds only what the panel already showed you: the store and the page's address without anything after a question mark or a hash, the product's name and price line, the verdict, the three answers, the shopper and agent rows, the fix's title and where to change it, and the buying test if you ran it. Never your email address, a phone number, a card number, a name, your install number or any text from the page.
It is kept for 90 days and then removed. You can delete it sooner: the browser that made the link keeps a private key for it, and Delete this link, in the panel's share menu, removes the page at once. We store that key only in a form that cannot be turned back into it.
The weekly email about a store you track
Anyone can track a store from the panel and get a short email once a week of what changed on it. You do not need to own the store or have an account. Nothing is sent unless you type an address and tick the consent line under it.
What it keeps:
- the email address you type, and a one-way hash of it
- the store's domain and the page you asked from
- the version of the consent line you ticked
- last week's public facts about the store, to compare with the next week
Every email has a one-click unsubscribe. Using it erases your address at once, and you can also ask us to erase it at hello@cresva.ai.
The AI shelf on a product page
On a product page, the panel can show where that product ranks when shopping assistants are asked about its category.
With browser view off, nothing from the page is read or sent for it unless you click Track. The panel says to turn on browser view to see the shelf. With browser view on, the extension asks Cresva for the product's shelf when you open Details after inspecting a product page.
What that request sends, and what Track and Untrack send:
- the page address, with the query string removed apart from variant and sku, and the page's canonical link
- the product's name, brand and category, exactly as the page's own structured data states them
- the extension's random install identifier, which we store only hashed
It sends no page text, no prices and nothing about you. The answer comes from a reading we already hold. Opening the panel asks no AI model anything.
What is kept, and for how long.
Each time the shelf is shown, we record the product page and the hashed install identifier, so we can count how often people come back to it. If you press Track this product, the same fields are recorded as a request, and a person confirms the shopping question before the product is measured. Both records are deleted after 90 days by the same nightly job as scans.
Untrack removes your request. The product stops being measured once nobody else tracks it; if someone else still does, the card says so. A tracked product page is a public page, not anything about you.
Signing in, and what that sends
Store mode never signs you in. Signing in is asked for when you press Analyse with Cresva on a Meta Ad Library ad, or the first time you open the panel on your own Meta Ads Manager or Google Ads, described below. You can use the store panel indefinitely without ever signing in.
Signing in registers a client, starts a pairing and opens a Cresva tab where you consent. Your browser keeps a one-time verifier and exchanges it for a token that is stored in the extension's local storage. Everything the panel does with that access is a lookup, with one exception you grant on the consent screen: saving an ad analysis you ask for, described under Analysing an ad below.
Once signed in, and only while the ads panel is open on one of your own ad accounts, the panel sends:
- Which brands your grant covers. The list is scoped on our side, never filtered in the browser.
- Campaign performance, pacing and anomalies for the brand you select, and a budget recommendation when you ask for one.
- The question you type, as you typed it, when you ask about a campaign, together with the campaign ids if you chose any. This is the one place the panel sends something you wrote rather than something it measured, and the question is also sent to OpenAI to be read, as listed under Which features use an AI model.
- The sign-in token Cresva issued, in a request header, with each of the above. No cookie is attached, and no sign-in to any store or ad platform is ever sent: Cresva's servers read your ad accounts through the connections you made in Cresva.
Nothing is written to an ad account.
When the panel proposes a budget change it opens the approvals queue in your dashboard and stops there. The change is made in the dashboard, by you. The panel has no ability to enact one.
Nothing from your own ad account's pages is ever sent. On an ads page the extension reads the address of the tab and never the page itself. The one page content ever sent is a single public Ad Library ad you choose to analyse.
Analysing an ad in the Meta Ad Library
After a scan, the panel offers a link that opens the store's name as a search in the Meta Ad Library. It is an ordinary link: the extension sends Meta nothing, and Meta sees the search when you open it. On Meta Ad Library pages only, the extension adds an Analyse with Cresva button to each ad. Nothing is read until you press it. When you do, the extension reads only that ad: its image or video, its text, title and caption, the advertiser's name and the domain shown on the card, its start date and its Ad Library ID. It sends them to Cresva under your account, and we analyse the ad and save the result to your Competitor Intel.
Analysing an ad needs a Cresva sign-in, asked for the first time you press the button, and each analysis counts as one run on your Cresva plan, the same as an analysis started from the Competitor Intel page. A press that reads no ad is not counted.
The extension fetches a still image itself, from Meta's image server, with no Facebook cookies and no referrer, and sends a smaller copy. A video is never downloaded in your browser: the extension sends its address, and Cresva's server downloads it from Meta's media server, takes four frames and deletes the file. If the extension cannot fetch an image, Cresva's server downloads that image the same way. Only the ad you pressed the button on is downloaded, and only from Meta's own servers.
The ad is read by an AI model. The image, or four frames from the video, is sent to OpenAI with the ad's text and the advertiser's name, never the video itself. What is sent, and to whom, is listed under Which features use an AI model below.
We store the analysis, not the image. The extension never reads your Facebook account, feed, messages, cookies or anything else on the page.
How long an analysis is kept. A saved analysis has no automatic expiry: it stays in your Competitor Intel history until it is deleted. To delete yours, write to us at the address below and we delete them. An analysis made for a brand in your dashboard is also deleted when that brand is deleted.
Which features use an AI model
A store scan never calls an AI model. These features do, each only when you are signed in and only when you use it:
- When you analyse an ad in the Meta Ad Library, the ad's image, or four frames taken from its video, its text and the advertiser's name are sent to OpenAI to produce the analysis.
- When you ask about a campaign in the signed-in ads panel, the question you type is sent to OpenAI to work out which figures to fetch. The answer itself is built from your data without a model.
When you start a store run, only public facts about the store are sent to AI models. Up to 60 of its public product titles and its product types go to Google (Gemini) to work out what it sells, and, when fixed rules cannot place its category, up to 40 product titles go to Anthropic (Claude), then OpenAI, then Google if neither answers. When its catalogue cannot be read, its homepage title and description, its live ad headlines, its name and its address are used instead. Ten shopping questions about what it sells are then sent to ChatGPT (OpenAI), Claude (Anthropic), Perplexity and Gemini (Google); they name the store only when nothing else says what it sells. The brand names found in the answers are sent to Anthropic, then OpenAI, then Google, to check which are brands. Nothing about you is sent. It needs no sign-in. See Store run.
Under OpenAI's data policy for its API, OpenAI does not use what we send to train its models, and it keeps it only for a limited period to check for abuse before deleting it.
Daily limits
Scans are capped per install per day. The counter is keyed to the random install identifier described under Anonymous by default, not to you, and it exists to stop one install crawling other people's storefronts without limit.
What is stored
One row is written per scan. It holds:
- the hashed install identifier, never the raw one
- the host, and the page address in the stripped form above
- the page kind
- the verdict and the individual check results, including the short evidence strings the panel shows you
- whether a browser view extract arrived and how many bytes it came to, which is a size and never any content
- the engine version, how long the scan took, and the timestamp
- where we could work them out from the store itself, a category and a shipping country, used only for the comparison line
A try to buy writes one row too: the hashed install identifier, the store's host and the product page address, which steps worked and where the walk stopped, the market it tested, and how long it took. Putting a competitor beside a store runs the same scan against the address you type, from our servers, and stores it the same way as any scan.
The page text extract is never stored and never logged.
The up to 2,048 character text sample is used inside one comparison on the server and then dropped. It is not written to the database, it does not appear in any log line, and it is not returned in the response. What is kept about it is a boolean and a byte count.
Be aware of the one thing that does cross over: the short evidence strings in the stored check results are quotations from your page, so a page title or a price the page displays can appear there. Those are the same strings the panel shows you on screen.
Caching.
A verdict computed from a page we fetched ourselves is kept for up to 24 hours, keyed by host. The panel reuses it only while it is less than an hour old, so a repeat scan of the same store within the hour costs nothing; after that the store is read again. A verdict computed with help from your browser extract is never cached, because it describes one person's browser session and must never be served to a second person.
Your IP address.
Used as a rate limiting key, held in a short lived counter that expires on its own. It is not written to the scan record.
Two numbers a day, about the service and not about you.
We count how many scans were answered from our cache and how many were freshly read, as two running totals per day. Those totals carry no host, no install identifier and no address: a date and a word. They tell us whether the cache is doing its job, and they expire after fourteen days.
Retention, and what we do with it
Two uses, and no third.
- Producing the report you asked for, in the panel, at the moment you asked for it.
- Aggregate readiness statistics. Stored verdicts are counted together so the panel can tell you where a store stands against others in the same category and country. That comparison is only shown once at least thirty comparable stores exist, and it is a count of bands and never a list of stores.
The panel also compares a store against the Cresva AI Shopper Index, which is a named population of direct-to-consumer stores we crawl and publish, not the web at large. That line is withheld until the population reaches 100 stores, and it names the population and its size in the sentence, so a percentage is never shown without saying what it is a percentage of.
We do not sell this data, and we do not share it with advertisers, data brokers or any third party for their own purposes.
A store scan makes no calls to any AI model. The analysis is code reading your page, not a model reading it, and there is therefore no third party model provider in this path and nothing to train on. The features that do use a model are listed under Which features use an AI model above.
That includes the panel's AI assistants section. What it shows is read from a weekly record of what each assistant answered for that category: the question was asked once, for the category, on a schedule of ours, and the panel reads the answer. Your scan does not ask an assistant anything, the store you are looking at is not sent to one, and opening the panel a hundred times asks nothing a hundred times. Where we have not asked for a category yet, the panel says it has not been asked rather than asking on your behalf. The store run is separate and does ask, as described under Store run.
How long we keep it.
90 days. A job runs every night and deletes every scan and try-to-buy record older than that. It is not a policy we apply by hand on request: it is a scheduled deletion, it records what it deleted each night, and the number above is read from the same constant the deletion uses, so this sentence cannot quietly drift away from what the code does.
You can also ask us to delete yours sooner. Write to us at the address below.
Anonymous by default
Store mode is anonymous. No account, no sign in, no email address and no connected store: the extension asks you for nothing and we do not know who you are. Signing in is a separate, optional step that turns on the ads panel and Ad Library analysis, and only then do we know who you are. Everything on this page describes store mode unless it says otherwise.
- The install identifier is a random UUID your browser generates once, at install, from the platform's cryptographic random source. It contains nothing about you or your machine.
- It is kept in the extension's local storage and deliberately not in Chrome's sync storage, so it does not replicate to your other machines.
- It is sent as a request header and stored only as a SHA-256 hash. The raw identifier is never written to our database.
- It is a rate limiting partition key and nothing else. Your browser mints it, so it proves nothing to us and cannot be used as a credential.
- Uninstalling the extension removes it. A fresh install mints a new one, unlinked to the old.
Everything the extension keeps in your browser.
This is the whole list, and it is generated from the list the code itself is checked against rather than written out here by hand. None of it is in Chrome's sync storage, so none of it reaches your other machines.
Kept until you uninstall (local storage for this extension):
installId: a random identifier your browser generates once, at install, used to partition rate limits and nothing else.includeRendered: whether Read this page in my browser is on, which it is unless you switch it off. It is the setting that decides if a page text sample is sent at all.sectionState: which sections of the panel you have folded away, so the panel opens the way you left it.hasScanned: whether a scan has ever completed on this install, which is what stops the panel showing a first-run walkthrough for ever.owned:<host>: one entry per store you have told the panel is yours, so it stops asking. The host is the only thing stored and it never leaves your browser.checkShare:<check>: one entry per check link you made: the link, when it expires, and the private key that lets you delete it. The key is sent to Cresva only when you press Delete this link.cresvaSession: your Cresva sign-in, as an opaque token and the moment it expires. It is written when you sign in, and read before the ads panel opens on one of your own ad dashboards and before an Ad Library ad is analysed. An expired one is treated as no session at all rather than as a valid one. It is absent until you sign in, and it is removed when you sign out or when Cresva no longer accepts it. It also holds a refresh token, so a sign-in is renewed in the background rather than asked for again every hour.cresvaOauthClient: the identifier this browser's copy of the extension registered itself under, so signing in a second time does not create a second registration. It is an identifier and not a secret: the extension is a public OAuth client and holds no client secret at all. It is absent until the first sign-in.adsHintDismissed: whether you have dismissed the one line at the bottom of the panel that mentions checking your ad numbers. It is a single true once you press dismiss, it is absent until then, and nothing about it leaves your browser.scanOnClick: whether clicking the Cresva icon scans the page straight away. It is set to true by your first press of Scan, which the panel says before you press it, and you can turn it off in the panel. Nothing about it leaves your browser.ratePrompt: how many of your inspections have finished, whether you have copied a fix, and whether the panel has asked you to rate the extension and what you pressed. It exists only so the panel asks once, after your second successful inspection or your first copied fix, and never again. Nothing about it leaves your browser.
Kept only until you close the browser (session storage, which is never written to disk):
grantContext: the tab you clicked the Cresva icon on, as a tab id and, when Chrome supplies it, that tab's address. It is what lets a scan act on the tab you actually clicked rather than whichever one you are looking at a minute later, it is kept in session storage so it is never written to disk, it is treated as expired after ten minutes, and it is gone when you close the browser.
Opening the panel, and removing the extension
- Opening the panel sends one request to cresva.ai, at most once a minute, so our servers are ready when you press Inspect. It carries nothing about the page, the store, your install or you, and no cookie.
- From version 6.0.0, the extension also tells cresva.ai when it is installed, when the panel is opened, when a page is checked and when a store run starts and finishes, so we can see where people stop. Each is the event's name and your random install identifier, which we keep only hashed, in a count for that day, for about four months. No address, no store, no page and nothing about you.
- From version 6.0.0, the page Chrome opens after install checks one sample product on allbirds.com, a public store, through cresva.ai, so you see a result before your first click. It sends the sample's address and your random install identifier, as any check does, and nothing about you.
- When you remove the extension, Chrome opens a page on cresva.ai that asks, optionally, why. The address carries the extension's version and nothing else.
- If you answer, we count the answer you chose and that version for the month. There is no free text, no install identifier and no account, and your network address is used only to limit how often answers can be sent, never stored with the answer.
Changes to this policy
If we change what a scan sends, what is stored, or how long we keep it, we will update this page and move the date at the top of it. Material changes will also be noted in the extension's Chrome Web Store listing, which is the surface that tells you an update has been installed.
Contact
Privacy questions, deletion requests and anything on this page you think is wrong: hello@cresva.ai.
Cresva, Inc., Delaware, United States. This policy supplements the general privacy policy at /privacy, which covers cresva.ai and the signed-in product. What the extension actually does is described at /extension.