PreferiumChapter 01 of 04
The engine fixes what it finds — every day.
88 automated health checks read the client site, the fixes are chosen, and only what passes verification stays live. Publishing follows the control level you set.
< 60 s
from fix to live
64
fix points in the HTML
2.4 s
thinking time per choice
1
rolled back today
The loop
The run. Five steps, and the fifth points back to the first.
A run that stops in a draft is not a run — it is a to-do list.
Find → fix → deploy → verify → roll backRUNNING
01
Find
88 automated health checks read the site and set the health score.
02
Fix
The change is written, inside the guardrails you set.
03
Deploy
Served in the HTML at the edge — no queue.
04
Verify
A real browser re-opens the page after title, meta, H1 and schema deploys.
05
Roll back
Automatically, if the page got worse.
The category
45 %
The category's own tools state that the agent handles around 45 % of the work list — and that nothing containing text is published before the client has reviewed it. The rest becomes drafts that wait.
Preferium AI Edge
100 %
On Full autopilot the engine runs the whole loop itself — from finding to rolling back. The entire technical work list, deployed without an approval queue, visible in the panel, and every item can be undone with one click.
The category's figures are its own, from public documentation read in August 2026.
1 of 4 · Find
What it looks for. 88 automated health checks, health score 0–1000.
A finding does not sit there as an alert: whatever is safe to fix automatically gets fixed — and every change is in the log.
Crawler access
An AI crawler is blocked, gets a 404, or never reaches a page that matters.
Structured data
A page a model would cite has no schema to cite from.
Page title and meta
Title or description is missing, empty, or does not say what the page answers.
Canonical and duplicates
Two addresses show the same thing, and neither is marked as the original.
Entity signals
Name, address and phone number differ between pages, so the model is unsure who the client is.
Freshness
A page that used to be cited has sat untouched and is losing ground.
Internal linking
A page no other page points to is a page the model rarely finds.
Indexing signals
Sitemap, robots and IndexNow disagree about what exists.
The checklist
The health checks, spread across thirteen areas.
“88 checks” is easy to claim. This is what they actually are — run on every page, every night, with findings that close themselves when the site changes.
AreaChecksWhat they look for
Page titles5Missing, too short, too long, duplicated across pages, more than one on the page.
Meta descriptions4Missing, too short, too long, duplicated across pages.
Headings7Missing or multiple main headings, skipped levels, heading identical to the title, repeated across pages, and length.
Content10Thin, duplicate, low quality, oversized page or HTML, missing language, broken resources, heavy scripts and stylesheets, insecure mixed content.
Links9No internal links, too many links, empty link text, empty addresses, broken internal and external links, orphan pages, invalid attributes, pages with only one incoming link.
Images2Missing alt text and oversized image files.
URLs3Double slashes, uppercase letters, overlong paths.
Canonical links5Missing, invalid, pointing at another page, pointing outside the domain, or more than one.
Structured data4No schema on the page, invalid schema, missing recommended fields, and schema Google itself rejected.
Sharing tags11Open Graph and X/Twitter: missing blocks, missing image, missing or invalid card type, and text that gets cut in the preview.
Indexability12Error pages, redirects, noindex in the page or the header, gaps in the sitemap, non-indexable pages listed in the sitemap, conflicting robots directives, nofollow, meta refresh, insecure redirect targets.
Hreflang8Missing return links, no default, missing self-reference, invalid language codes, ambiguous regional codes, duplicates, broken targets, lone alternates.
Domain level8HTTPS, serving of sitemap and robots, unclassified site type, expiring certificate, and AI crawlers blocked for search, training or mixed purposes.
In total88Plus 24 extra technical checks from an independent second engine, and a full external crawl that hunts duplicates, redirect loops and non-indexable pages.
The audit
What the health checks give you back.
The checks are not an alert list that only grows. Findings open and close themselves as the site changes — and the audit gives you five views that answer what is wrong, where, and what should be done first.
ScoreHealth score 0–1000One score for the whole site, with the week’s change and what moved it. The list also shows what the site already passes, not just what it fails.
MapStructure tree and page explorerThe whole site as a navigable hierarchy, plus every page in a sortable table with score, status and focus keyword.
GapsOrphan pages and crawl depthPages nothing links to, and pages buried so deep they are invisible in practice. Internal link suggestions are found by meaning, not word match.
PruningContent pruning reportEvery page sorted into keep, thin or prune — judged against actual Google impressions instead of a guess. Reversible with one click.
AdvisoryReadability and brand voiceReadability score, sentence length and how closely a page hits the voice you defined. Advisory only — never published to the live page.
Disagree?Ignore a findingHide a check for one page or the whole site without skewing the score. Plus an independent second audit with 24 extra technical checks.
OrderRecommendation feedA prioritized list of what to do next, scored daily by likely impact — so the most valuable item sits on top instead of the newest.
PacePublishing paceHow often the client publishes, and how many new pages arrive each month. Makes changes in the site’s publishing rhythm visible over time.
TypeSite type detectionThe engine works out what kind of site the client runs — store, blog, local business, publisher — and picks the structured data that fits that type.
Help that knows the numbers
It can read everything, and change nothing.
The assistant in the panel answers from the site's actual numbers — not from generic SEO advice pulled off the web.
SEO assistantA conversation that knows the site’s actual numbers and answers with concrete, linked steps.
Action plansAsk for a plan and get ordered steps, each linking straight to the screen where the work is done.
Queue for what comes nextWhat the automation intends to work on next, and why, before it runs.
Command paletteJump to any page, site or setting from a keyboard search.
2 of 4 · Fix
What it fixes. Item by item, in the HTML at the edge.
The health checks find — these items are the most common places the engine fixes. The complete open list of 358 capabilities across 24 areas lives in the capability catalog: 328 capabilities in 23 areas are client-facing. 30 capabilities in 1 area are agency and white-label functions.
64items in total
Technical foundation
What search engines and models read first. Fixed in the HTML that is served, without anyone changing how the client publishes.
01Title tags and meta descriptionsThe blue headline in Google and the strongest signal about what a page is about. Unique, compelling titles per page for a higher click-through rate.
02Image alt text and social metadataGoogle cannot see images, so alt text that explains the subject is added — and images are compressed so they load fast on mobile.
03Core Web Vitals and speedGoogle’s grading system for user experience. Loading (LCP), interactivity (INP) and visual stability (CLS) are optimized.
04Mobile friendlinessGoogle looks at the mobile version first. Readable text, tappable buttons, and content that adapts to small screens.
05Sitemap.xml and robots.txtThe map and the rulebook the engines get: where the pages are, and where not to look.
06Canonical and paginationSays which page is the original, so variants of the same product are not read as duplicate content.
07Schema.org / structured dataThe language machines speak. Price, stock status and ratings are marked up so they can show directly in the result.
08Breadcrumb markupThe breadcrumb trail is marked up, so the engine can show the path to the page instead of just the URL.
09Heading structure and H1One main heading per page and logical subheadings, so the text reads well for both people and robots.
10Redirects and error handlingWhen a page is deleted or moved, a 301 redirect to the nearest relevant page is set up automatically.
11Noindex/nofollow where it belongsFilter views, search result pages and privacy pages are kept out, so the effort goes to what sells.
12HTML cleanup and validationCode errors such as missing closing tags are found and fixed, so the page reads clean to the engines.
13HTTPS and HTTP/2Security is a ranking signal. The lock is checked, and so is a modern transport.
14Hreflang for multilingual sitesThe right language version to the right country, so a Swedish visitor does not land on the German page.
15Indexing and re-indexingNo waiting for the engine to drop by — APIs report new or changed content the moment it ships.
16Open Graph and X/Twitter cardsHow the page reads when shared on Facebook, LinkedIn and X: title, description, image and card type — added where missing. An image address that might not exist is never invented.
17Security headersModern security headers are added to every response served, without touching the client’s server setup.
18Home-screen and app iconsTouch icon, tile color and theme color, so the site looks right when someone saves it to their phone.
19SEO plugin markup set asideRank Math, Yoast, All in One SEO and Squirrly are detected and their competing schema is replaced by one merged graph, so machines do not get two different answers.
20Schema coverage reportWhich valuable schema types the site still lacks, page by page — out of 31 types the engine can write.
21URL hygieneDouble slashes, uppercase letters and overlong paths are flagged before they become an indexing problem.
22Broken links and resourcesInternal links, outbound links, scripts, stylesheets and images are tested for errors and reported, so you decide what gets redirected.
23410 for retired pagesMark a page permanently gone, so search engines and AI crawlers drop it — instead of redirecting it forever.
24Organization identityLegal name, logo and official profile links marked up and tied to the Wikidata identifier, so machines can verify who the client actually is.
25Author profilesReal people credited on articles with role, bio and verified profiles, so machines see who wrote what.
Keywords and pages
Which queries the client can actually win, and which page should answer each of them.
01Keyword researchNo guessing. Large datasets show exactly which words customers use, including misspellings and local variants.
02Primary keyword per pageOne clear lead term per page, placed in title, URL and heading without looking forced.
03Semantic coverage and LSIRelated terms that make the topic unambiguous — write about Apple, and iPhone and iOS help the engine know which Apple.
04Long tail and question queriesFewer searches per phrase, but visitors who are ready to buy.
05Keyword mapping to pagesEvery keyword gets its own home, so five pages do not fight over the same term.
06Gap analysis against competitorsThe client site laid over a competitor’s to see where content is missing.
07Competitor analysisWhich terms do they rank for that the client lacks? The holes in their strategy become the client’s advantage.
08Internal anchor textLink text tells the engine what the target page is about. “Read about running shoes” beats “click here”.
09Ongoing SERP monitoringThe result page changes — map packs, videos, images. The content is adapted to the new picture.
10Movement reportWhich terms bring the most traffic, and which are on the way up. The cold facts.
Content and structure
How the pages connect, and how the models understand what belongs to what.
01Internal link structureLinks are roads. Smart roads between pages show which ones matter most.
02FAQ and HowTo snippets where they fitQuestions and answers formatted so the engine can show the answer itself in the result.
03Product and category copyThin descriptions rank poorly. Copy can be extended with buying help, benefits and use cases.
04Duplicate controlPages that are too alike are found and merged, or one is marked as the one that counts.
05Canonical for variantsThe same product in several colors or sizes points to one original.
06Alerts on breakageThe smoke detector: if traffic drops or an important page disappears, you hear about it immediately.
07Navigation is never touchedVisible rewrites deliberately skip everything in the menu, footer and sidebar.
08Decorative images are left alonePurely decorative images are recognized and skipped — no description is written or forced.
09Right-to-left languagesCorrect text direction for Arabic, Hebrew and similar languages, also in text the engine inserts.
10Map box from coordinatesMap boxes for physical addresses are built safely from coordinates instead of pasted third-party code.
Local visibility and authority
The signals that tie the client’s name to a place and a field — and to other sources that mention them.
01Structured NAPName, address, phone — written exactly the same everywhere, so the engine is never in doubt about who the client is.
02Google Business Profile alignmentOpening hours, address and categories compared with what the site publishes; mismatches are flagged for you.
03Knowledge Graph signalsThe goal is that the engine understands the business as an entity, not just a website.
04Link analysisWho links to the client? Toxic links from spam sites are surfaced, good opportunities are found.
05Citations and mentionsListings in relevant directories and registers act as votes that the business is real.
Measurement
Reporting against GA4 and Search Console, and on who the models cite — with the numbers that actually moved.
01Monthly reportOnce a month the report sums up what was changed and what moved, per client — white-label, ready to forward.
02GA4 and Search ConsoleMeasured against the client’s own accounts, not invented numbers. Everything reported can be checked.
03Citations in AI answersTracked prompts are run against the engines every week, in the client’s market, and the answer is read: who is cited.
04Which AI crawlers visitedRead straight from the request at the edge, not from a tracking script — a crawler runs none.
05Impressions, clicks and keywordsHow often the client shows up, how often someone clicks, and on which queries.
06Movement per pageEvery page is followed on its own, so you see which ones lift and which ones stand still.
07What changed, and what workedThe report separates changes that were made from changes that actually moved a number.
Control and insight
The engine works on its own; control stays with you the whole way, in real time.
01Full log of every changeEvery change listed with its time, the page it applied to and what was changed.
02See what is live, any timeThe panel shows the current state at all times — not yesterday’s report.
03Withdraw a change with one clickDisagree? Roll it back yourself. The previous state is served again immediately.
04Adjust a change yourselfWant to keep the change but phrase it differently? Edit it in the panel.
05Switch to review when you wantWant to look before anything goes live? Set the client to Review everything, and every change waits for your approval.
06No plugin in the CMSThe source code in the publishing system is never modified. Everything is delivered in the HTML response at the edge.
07No visible widget for visitorsThe work happens in the machine room the engines read. The design visitors see stays exactly as the client wants it.
Autopilot
How autonomous is it, really?
“Autonomous” is easy to write and hard to stand behind. This is what the word means here: what happens without anyone pressing a button, what waits for you, and what never happens at all.
Level 1DefaultFull autopilotThe engine deploys on its own and checks afterwards that the page actually got better.Deployed and verified
Level 2Most common at the startInvisible-onlySignals only machines read (title, meta, schema) go straight out. Text people see on the page waits for you.Deployed — this was technical
Level 3Full brakeReview everythingNothing is deployed until someone on your team has seen the proposal and clicked yes.Waiting for you
Around the clockWithout you lifting a finger
Nightly runThe whole site is re-read every day. New pages are picked up without anyone reporting them.
Self-repair every 30 minutesThe live page is compared with what it should be. If something overwrote the work, it is put back.
One-click fixes waiting for youWhat the engine does not do on its own is collected as clear buttons you can review and press when you are ready.
Recommendation feedNext steps are scored daily by likely impact, so the most valuable item sits on top instead of the newest.
Live in 60 secondsA change you make in the panel is out at the edge within the minute. No republish, no queue.
The limit
Some things it never does on its own.
Not because it cannot, but because the cost of getting it wrong is too high for a machine to make the call alone.
Risky fixes are assessed firstRemoving a noindex or a conflicting canonical goes through a safety assessment and is flagged for you.
Visible text is off by defaultAnything a visitor can see is off until you deliberately turn it on — even when the autopilot is otherwise full.
Menu, footer and sidebarNavigation is never touched, at any level. It is the client’s structure, not a signal to optimize.
3 of 4 · Fix
What a change looks like.
On the left is the HTML as it often is. On the right is the same page with the signals in place — the ones search engines and language models actually read.
northwindcycles.com/bike-service6 changes
Before
<title>Home | Northwind Cycles</title>
<meta name="description" content="">
<img src="workshop.jpg">
After
<title>Bike Service in Seattle — Book Today</title>
<meta name="description" content="Service for every bike. Reply within one day.">
<link rel="canonical" href="https://northwindcycles.com/bike-service">
<meta property="og:image" content="/og-workshop.png">
<img src="workshop.jpg" alt="Bike service in the workshop">
<script type="application/ld+json">{"@type":"Service"}</script>
4 of 4 · Deploy
Why the engine is built this way.
Everything rests on one fact: AI crawlers do not run JavaScript. They read the HTML as it stands. Three ways to deliver a change — follow the request through each of them.
GPTBot, ClaudeBot, PerplexityBotJavaScript: off
Script on the pageMisses
The change runs in the browser
A script is added to every page and rewrites the content when a browser runs it.
The crawler never reaches the change. It reads the HTML before the script has run — and does not run it at all.
Bot detectionFragile
The crawler gets its own version
The system recognizes which crawler is asking and serves it a separate, pre-rendered HTML.
Two different answers for the same page. The list has to be kept current — a crawler that is not on it falls outside until someone adds it.
Rewriting at the edgeOurs
The HTML is changed before it is sent
The HTML is rewritten in the response itself. One CNAME record, no plugin, nothing added to the pages.
One answer for everyone. On a server-rendered page there are no other versions, so there is no detection that can miss.
The comparison is from August 2026 and is based on the vendors' own public documentation. The claim that crawler and human get exactly the same HTML applies to server-rendered sites; if the site is pure JavaScript, dynamic rendering is used.
The shape assistants cite
A model prefers to cite a finished answer.
So the engine can place an answer block at the top of the page — in exactly the form assistants prefer to pull from. It stays off until you enable visible rewrites.
What the block looks like
Short summaryOne or two sentences that say what the page answers — the first thing a model reads.
Direct answer paragraphA paragraph written as an actual answer, not as sales copy.
Key takeawaysThree to five points an assistant can lift verbatim without rewriting them.
SourcesWhere the claims come from, so the answer can be checked — by the model too.
Labels are written in the page's own language, in 21 languages. Content we insert gets the correct text direction for Arabic and Hebrew as well.
Related links boxUp to seven internal links chosen by meaning, not word match — helps both readers and crawlers find related pages.
Map box from coordinatesFor physical addresses, built safely from coordinates instead of pasted third-party code.
List items and table textBullet points are rewritten to be complete, and tables get a caption and column headers machines can understand.
Code explanations and captionsA short plain-language explanation above code examples, and captions that say what the image actually shows.
Form labels and link textClear labels and helpful placeholder text — and “click here” becomes text that says where the link goes.
Duplicate titles avoided while writingThe titles of nearby pages are shown to the AI as it writes, so duplicates are avoided immediately instead of being found in an audit weeks later.
The files
Ten files we generate and serve for you.
The files search engines and AI agents look for, generated from the live site and kept current. No plugin, no file to maintain.
FileWhat it doesServed by
/sitemap.xmlEvery page with sensible priorities; blocked and non-indexable pages are left out.Platform
/robots.txtAI crawlers are allowed by default, and the client’s own rules are respected.Platform
/llms.txtThe new standard file that gives language models a guided tour — in short and full versions.Platform
/agent.jsonMachine-readable manifest describing the site for AI agents.Platform
/agents.txtThe same description in text form, for agents that look for it.Platform
Content signalA clear declaration of what AI may do: appear in search, answer questions, train — your choice.You
/feed.xml · AtomContent feed generated from the client’s pages.Platform
Web app manifestSo the site behaves like an app when saved to a phone.Platform
/.well-known/security.txtPublished security contact, so researchers can reach the client the right way.Platform
IndexNowSearch engines are notified the moment a page changes, instead of waiting for the next crawl.Platform
Priority and update frequency can be overridden per page, and thin or private pages are kept out of the map you hand to search engines entirely.
In the same run
Speed and language are signals too.
A page that loads slowly or answers in the wrong language gets cited less often. Both are measured on real visitors, not in a lab simulation.
Speed and delivery
Numbers from real Chrome visitors.
Core Web VitalsLoading, responsiveness and visual stability, taken from real Chrome visitors.
PageSpeed auditsGoogle’s own audit of performance, SEO, accessibility and best practices — automatic and on demand.
Critical CSS and preloadThe styles for the first screen are inlined, and the files that matter most load first.
AVIF and WebPPicture markup so the browser can fetch a much smaller version, with a safe fallback to the original.
Lazy loadingImages and embeds below the fold load only when the visitor scrolls to them.
Stays up when the origin goes downOn a timeout or error, the last good copy keeps being served, so an outage does not take the site down.
Language and international
The right version to the right visitor.
Hreflang written and servedThe links between country and language versions are delivered, so the right version reaches the right visitor.
Eight hreflang checksFinds missing return links, invalid language codes, ambiguous regional codes and broken targets.
Translation with grammarTitles, meta and headings translated into the languages you want, with correct regional grammar.
Glossary that locks wordsProduct names, brand terms and technical vocabulary that must never be translated.
Serving per languageEvery visitor and crawler gets the right version, with a sensible fallback when there is no exact match.
Inserted text in the page’s languageAnswer block, links box and map labels write their own labels in 21 languages.
Want to see it on a client site?
Pick one of your client sites. We run it through the health checks, show you what the engine would fix first, and walk you through the panel your clients would see under your brand.
Agency agreementNo lock-in on your dataEvery change exportablePricing on the pricing page