Preferium AI Edge · White label for agencies

Autonomous AI-SEO your agency resells under its own brand.

Your clients connect a site with one DNS record. The engine finds, fixes, deploys and verifies at the edge — and gets them cited by ChatGPT, Claude, Perplexity and Gemini. Your logo, your Stripe, your prices.

Join the waitlist See how it works

Readable for

Google
ChatGPT
Perplexity
Gemini
Claude
Grok
Copilot
Siri

northwind-cycles.exampleEdgeautonomous run16%

Optimize northwind-cycles.example for AI search.

Thinking … 2.4 s · 88 health checks × 12 pages

READread_page/48 signals

read_page×11 more pages214 signals in total

findall checks run on 12 pages18 findings

SELECTTechnical foundationcheck[7 found]

Keywords and schemaanalysis[6 found]

Internal linkingstructure[5 found]

6 of 18 findings chosen now — 12 queued for tonight's run

CHANGE/index.html + 5 pagesat the edge, never in the CMS

<head>

- <title>Home</title>

+ <title>Bike Service & Repair in Seattle …

+ <meta name="description" content="…

+ <script type="application/ld+json">

+ <link rel="canonical" href="https://…

+ 2 more: alt text (l. 88) and og:image (l. 45)

6 changes written — all reversible with one click

VERIFYReal browser + GPTBot fetch again6 of 6[ok]

Live at the edge · 0 rolled back · next run 03:10

Control level

Full autopilotInvisible-onlyReview everything

The level applies to the whole site and can be changed at any time — even mid-run.

See the three levels

Last run07:28

6changes deployed12pages in the run0rolled back58 sfrom finding to live

Citation share+4

31%

Live changesnow

title/pricingUndo

schema/aboutEdit

metahome pageUndo

alt text/imagesEdit

canonical/blogUndo

h1/contactEdit

og:imagehome pageUndo

sitemapallEdit

title/pricingUndo

schema/aboutEdit

metahome pageUndo

alt text/imagesEdit

canonical/blogUndo

h1/contactEdit

og:imagehome pageUndo

sitemapallEdit

Sample data · Northwind Cycles

Product

One meaningful state per card.

Six parts of the engine, laid out the way your clients meet them in the panel. The numbers are sample data from a fictional client, Northwind Cycles.

Preferium AI Edge/ Six of 328 client-facing capabilities

Autonomous optimization

On Full autopilot the engine finds, fixes, deploys and verifies on its own — inside the guardrails you set, with every change reversible.

AI visibility

Tracked prompts in ChatGPT, Claude, Perplexity and Gemini.

12,480+18.4%

Impressions, last 30 days

Agent traffic

Crawler visits and AI referrals, kept separate from the answers.

ChatGPT/featuresCited

Claude/aboutRead

Perplexity/pricingCited

Gemini/blogRead

AI Overviews/glossaryCited

Control and safety

Approval, kill switch, undeploy and an honest CMS boundary.

Approval per change if you want it

Kill switch and undeploy

Scope control: what it may never touch

Pause and work limits per site

Everything logged in the panel

Verification

A real browser opens the page again, and rolls back if it got worse.

✓Deployed at the edge07:31

✓Re-fetched07:32

✓Compared07:33

opening the page again …

Edge delivery

Changes are delivered reversibly in the HTML response, not in the CMS or the source code.

Visible changes3 tags

Invisible signalsschema

Reversible100%

Capability catalogSix cards here. 358 capabilities in the full catalog.31 schema types, four AI assistants with prompt tracking, and every switch you control the work with. 328 capabilities across 23 areas are client-facing; 30 capabilities in one dedicated agency area are for your agency.Meta tags 15Structured data 12Audit 33AI visibility 26Control 42Security 16

Scope

24 areas. 23 for client sites. One for the agency.

The whole engine, grouped. 328 client-facing capabilities across 23 areas, plus the agency and white-label area that only you see.

Capability catalog358 capabilities · 24 areas

Open the catalog and search it328 capabilities across 23 areas are client-facing. 30 capabilities in the agency & white-label area are yours.358 capabilities

The engine

One run. Analyze, prioritize, deliver, verify.

Five steps in a fixed order, not decoration — the engine runs the whole loop every day, whether or not anyone opens the panel.

The autonomous run

Five steps, run in order, every day.

RUNNING

READ
Reads
The whole site is re-crawled every day — new pages are picked up without anyone reporting them.
PICK
Prioritizes
88 automated health checks run per page. Findings are ranked by likely effect, not by age.
FIX
Changes
The fix is written into the HTML that is served, in front of the site — never inside the client’s CMS.
CHECK
Verifies
Title, meta description, H1 and schema deploys are re-opened in a real browser afterwards, and fetched as GPTBot, to confirm the change is really there.
UNDO
Rolls back
If a page got worse, the change is reverted automatically — and you are told that it happened.

Signals

Citations, mentions and sources are not the same thing.

A citation is the client’s own page used as a source. A mention only names them. Sources are the domains the answer was built on. Three measurements from the same tracked answers, kept apart — here side by side over twelve weeks. AI referral visits and crawler fetches are counted separately, under Agent traffic.

Visibility over time

Trend without claimed causality. The curves can move independently.

12 weeks · sample data

Week 1Week 3Week 5Week 7Week 9Week 12

Citation share
31%+4
How often a tracked answer uses one of the client’s own pages as a source. Sampled prompts on a schedule — not access to anyone’s real conversations.
Share of voice
58%+6
How often the client is named against other brands in the same answer. Being named is not being cited, and it is not a visit.
Sources behind the answer
46+9
How many domains the tracked answers built on over twelve weeks — the client’s own and everyone else’s.

Citation share

31%+4

in tracked answers

Share of voice

58%+6

naming the client

Sources behind the answer

46+9

domains in 12 weeks

Visibility trend

12 weeks

no causality claimed

Tracked answers

What the engines answer when someone asks about your client.

We run ten selected prompts per page on a schedule and read which page gets cited, and in which position. Synthetic method — no access to anyone's real conversations.

Tracked answers

Where the client stands in selected prompts, per engine.

Sample data

Google/features#1
ChatGPT/features#1
Perplexity/about#3
Gemini/blog#5
Claude/features#2
AI Mode/contact#8

Synthetic method — we run selected prompts on a schedule and never read anyone else's conversations.

Prompts mapped per page

Ten prompts per page, and the sources that carry the answer.

10 prompts

The prompts we run

ai seovisibilitychatgptschemakeywordsedgeperplexitygeministructured datalocal

Sources behind the answer

the clientwikipediatrade pressreviewsown siteforumeditorialowneddirectory

10prompts mapped per page9sources behind the answerSample data

Agent traffic

Who fetches the pages, and what they feed.

Answer crawlers and training crawlers are not the same. The log keeps them apart, so you see which fetch can actually become a citation.

Crawlers fetching the pages

Every agent in the log, and the engine it feeds.

6 agents · 7 days

OAI-SearchBotOpenAIChatGPTAnswer82%2522 h ago
PerplexityBotPerplexityPerplexityAnswer64%1991 h ago
Claude-UserAnthropicClaudeAnswer41%1275 h ago
GooglebotGoogleAI OverviewsAnswer39%12330 min ago
GPTBotOpenAIChatGPTTraining55%1693 h ago
Google-ExtendedGoogleAI OverviewsTraining48%1486 h ago

New crawlers appear as vendors roll them out. If one is blocked, the panel shows which pages it could not read. Sample data.

AI referral traffic

A visit from an AI tool is not a citation.

A separate signal. We measure where they land and how far in they go. Sample data.

5pages linked togetherInternallink structure

Click depth1.8clicks on average−1.2Landing page64%share on the top entry pageContinues41%to a third page+9

Delivery

Follow one change from finding to verified.

48 signals are read, six changes are chosen, and only what passes verification stays. The rest is rolled back automatically. Sample data.

From signal to delivered

Where the 48 signals stand today.

Last run

Signals read48 · 100%

Findings with an action18 · 38%−30 within guardrails

Changes selected6 · 13%−12 low effect

Deployed and verified5 · 10%−1 rolled back

2.4 s

thinking time before the pick

pages in the run

elements in the HTML

rolled back today

One change, opened up

title · /

Verified

Found07:28<title> empty on 3 of 12 pages.
Written07:30New title proposed within the guardrails you set.
Deployed07:31Rolled out in the HTML at the edge, not in the CMS.
Verified07:33A real browser re-opened the page and approved it.
Logged07:33Visible in the panel, reversible with one click.

Control

Who presses publish? That is the whole choice.

The engine finds the same improvement whatever you choose. The level only decides what happens to it next: it ships on its own (Full autopilot), it ships only if it is a signal that machines read (Invisible-only), or it waits for someone to approve it (Review everything). Below, one real change follows all three paths.

Step 1 · the proposal at 07:28/index.htmlfield title · 64 → 51 chars

propose title · home pageHome - Northwind Cycles Bike Shop Seattle WA Bikes Parts RepairsBike Service & Repair in Seattle | Northwind CyclesThe title is shorter, clearer and names what people actually search for. This exact proposal is the starting point in all three columns below — only what happens next differs. Sample data.

Step 2 · Three paths forwardSame proposal, three levels. Follow each column top to bottom.

Level 1Default

Full autopilot

The engine deploys on its own and checks afterwards that the page actually got better.

Found at 07:28
Quality gate passed
Deployed at the edge
Verified in a real browser

This change, at 07:31Deployed and verifiedNobody waited for anyone. The new title was live three minutes after the engine found the gap.

Level 2Most common at the start

Invisible-only

Signals only machines read (title, meta, schema) go straight out. Text people see on the page waits for you.

Found at 07:28
Classified: technical signal
Deployed at the edge
Visible text would have waited here

This change, at 07:31Deployed — this was technicalA title shows in Google, not on the page — so it shipped. Had the proposal been a heading people read, it would have waited.

Level 3Full brake

Review everything

Nothing is deployed until someone on your team has seen the proposal and clicked yes.

Found at 07:28
Added to the approval queue
Waiting for approval
Not deployed

This change, at 07:31Waiting for youIt sits in the queue with before/after and a reason. One click deploys, one click discards — nothing happens by itself.

The level is set once in the panel and applies to the whole site. You can change it at any time — even mid-run — and pick a stricter level for individual pages.See the full control chapter

Safe delivery

Everything can be undone.

Changes are layered onto the HTML response, not written into the client's site. So there is always a way back — whatever level you run on.

Undo one changeFind the change in the panel and pull it back. The page answers as before within seconds.

Undo everything at onceDisconnect from the edge and the site falls back to its original HTML — all of it, at the same time.

The CMS is never touchedThe engine does not write to WordPress, Webflow or the source code. Everything there stays as it was.

Page coverage

Every URL, and what happens next.

One row per URL: whether the AI engines have fetched the page, whether they cite it, and what the engine does with it next. A page that has never been fetched does not exist in the answers — however good the copy is. Sample data.

Coverage · 12 pages

Cited3The AI engines use the page as a source in answers.Kept fresh

Fetched4A crawler has read the page, but it is not cited yet.Sharpened for answers

Not fetched3No AI crawler has been by. The page does not exist in the answers.Pushed via IndexNow

Blocked2The site shuts the crawler out, or the page answers with an error.Fixed in robots

Page coverage

Every URL, the agents that fetched it, and what happens next.

Control & safety/features/control-and-safetyNever fetched—Published three weeks ago, never fetched by an AI agent. Invisible in answers today.Submit URL
Agent traffic/features/agent-trafficFetch errorChatGPT404Answers 404 to GPTBot. The agent sees nothing at this URL.Inspect
Glossary/glossaryAwaiting re-fetchClaudeChatGPT200Changed two days ago, last fetched five days ago. The agents are reading old text.Request re-fetch
Pricing/pricingNo answer fetchChatGPT200Only training agents have fetched the page, so it cannot be cited yet.Ping crawlers
Autonomous optimization/features/autonomous-optimizationFreshChatGPTPerplexityGoogle200Fetched daily and cited in ChatGPT. Healthy.Open
AI visibility/features/ai-visibilityFreshPerplexityGemini200Cited in Perplexity since 14 June.Open

How it works

Set it up in the panel. Then it runs.

All setup happens in the panel, once per client: control level, guardrails, work limits and alerts. From there the engine runs on its own — you open the panel because you want to, not because it is waiting for you.

01Connect a clientOne DNS record per client site. The engine detects the platform, fills in the brand details and reads the site as it stands: technical foundation, keywords, structure and what the models see.
02Set the guardrails — onceIn your branded panel: control level, what it may never touch, work limits per site and which alerts you want. That is the whole setup.
03Autonomous operation from day oneThe engine finds, fixes, deploys and verifies on its own, every night, without anyone approving. You report to the client under your brand, with GA4 and Search Console beside the AI numbers.

One DNS record at the clientPropagated

Type: CNAME
Name: www
Value: edge.preferium.com

That is the whole connection. No plugin, no code on the pages, no access to the client's server.

Autonomous work planRunning

Read the siteHealth checks on 12 pages, health score set
Pick the changesSix of eighteen findings, within guardrails
Deploy at the edgeRolled out in the HTML response, no queue
Verify and logA real browser re-opens the page

White label

Your platform. Your brand, your Stripe, your prices.

For agencies

Agency partner

Private preview

You run Preferium AI Edge for your clients under your own brand. Connect your Stripe, build your packages, set your prices — every client payment runs on your Stripe and the margin is yours.

Your logo, your colors, your URL — clients only ever see your brand
Your own Stripe: clients pay you, you keep the margin
Three control levels: Full autopilot, Invisible-only and Review everything
Changes delivered reversibly at the edge, never written into the CMS

Join the waitlist

Registration opens to founding partners first. Live plan pricing is on the pricing page.

At scale

Enterprise

Talk to us

For established agencies and networks: a branded landing page on your own domain with signup, more clients, more seats, audit-logged and isolated.

Branded landing page on your own domain with signup
More clients and seats, isolated per client and audit-logged
Same engine, same three control levels

Talk to us

Scope and onboarding agreed with you.

Compare plans on the pricing page

Questions

In short.

The questions agencies ask most. The rest is in the terms, in full.

Read the terms

What is Preferium AI Edge?

An autonomous AI-SEO platform built by Preferium AS that agencies resell under their own brand. It connects to a client site through one DNS record, optimizes the HTML at the edge, and tracks who ChatGPT, Claude, Perplexity and Gemini cite. Your clients see your agency, not Preferium.

What does “autonomous” mean here, concretely?

That the engine finds, fixes, deploys, verifies and rolls back — on its own, every day, within the guardrails you set. With the switches on, nobody has to open the panel: everything is there when you want to look, and every change can be undone with one click.

Does it deploy without anyone reviewing first?

That depends on the control level. Full autopilot deploys visible changes too. Invisible-only ships only the signals machines read — title, meta, schema — and holds visible text for review. Review everything waits for approval on every change. You pick the level per client in the panel.

Join the founding-partner waitlist

Partner registration is not open yet. Leave your details and we will invite you as we open it — or talk to us about Enterprise.