Most guides treat this as two jobs. One article tells you how to measure SEO performance and another tells you how to run an audit, and the reader ends up doing them in the wrong order. You audit, you fix a list of things, and then you go looking for a number that proves it worked. By that point the number you needed was the one you did not write down before you started.
They are one procedure. The audit tells you what to change and the measurement tells you whether changing it mattered, and the measurement has to be set up first or it cannot answer the question. What follows is that procedure in order, with the parts that come from Google’s own documentation quoted so you can check them rather than take our word for it.
What SEO performance is, and why the audit cannot come first
SEO performance is not a score. It is the set of things that happen when somebody searches, which is how often you were shown, how often you were chosen, and what those people did next. Every tool that hands you a number out of 100 is summarizing a page, not measuring any of that.
That distinction sets the order of work. If performance is a set of observed events, then the only way to know whether a fix worked is to have a record of those events from before the fix. An audit run without that record produces a list of improvements and no way to price any of them afterward.
So the procedure has five stages and the first one is not the audit. Capture the baseline. Run the audit passes. Order the findings. Set the re-measure date. Compare. Teams that skip stage one are the ones still arguing in month four about whether the work was worth it, because both sides are describing a period nobody recorded.
This sits inside the broader diagnostic work our SEO audit page sets out, and the rest of this article is the run book version of it, written so you can do it yourself on a Tuesday.
Stage one, capture the baseline before you touch anything
This takes about twenty minutes and it is the step with the highest cost of skipping, because it is the only one that cannot be done later.
Export the last full period of search data before any change lands. Not a screenshot of a dashboard, an export you can open again in six months when the dashboard has moved on and the comparison window has rolled past the date you care about. Put it somewhere shared, with the date in the filename, and note what shipped that week so you can tell your own deploys apart from everybody else’s.
Capture the same export at the page level as well as the site level. Site totals hide the thing you usually need, which is that two page groups moved in opposite directions and canceled out. A site-level line that looks flat for a quarter is very often a winning section and a losing one of roughly equal size, and you cannot separate them after the fact from a total.
One practical warning if the property is new. Google notes that “it can take up to a week to generate data for a newly created site, or a site newly added to Search Console”, so if you have just connected it, your baseline is not ready yet and waiting is the correct move.
What to export, and the metric that will mislead your baseline
Four metrics come out of the search performance report and Google defines each of them plainly. Getting one of them wrong at baseline stage poisons everything downstream.
Impressions are “how often someone saw a link to your site on Google”, and Google adds that “depending on the result type, the link might need to be scrolled or expanded into view”. Clicks are “how often someone clicked a link from Google to your site”. Click-through rate is “the calculation of (clicks ÷ impressions)”. Those three are counts and they behave themselves.
The fourth does not. Average position is “the topmost position occupied by a link to your property or page in search results, averaged across all queries in which your property appeared”, and the sentence that matters for your baseline comes just after it. “A link must get an impression for its position to be recorded.”
Read that as a procedural instruction rather than as a definition. Queries where you never appeared are not in the average at all, so the biggest win an audit can produce, pages starting to appear for things they never appeared for, arrives as new rows rather than as movement in existing ones. Export the row count and the total impressions alongside the averages, or the improvement will be invisible in the only column you were watching. What each of these numbers actually means once you have them is a separate subject, and our explainer on what your SEO ranking actually is takes the position figure apart properly.
Stage two, pass one, can Google reach the pages at all
Now the audit, and it runs in three passes rather than as one long checklist, because the passes have a dependency order and doing them out of order wastes the work.
Pass one asks a single question about every page that matters. Can a search engine get to it, and does it see what a person sees. Google puts the standard directly, saying that “when Google crawls a page, it should ideally see the page the same way an average user does”, and that it “needs to be able to access the same resources as the user’s browser”.
In practice that is a short list. Which of your pages are known and which are not. Whether anything important is blocked. Whether pages return the status code you think they do rather than a soft failure that looks fine in a browser. Whether content that only exists after scripts run is visible at all. And whether redirects land where they claim to.
Everything in later passes assumes this pass came back clean. A perfectly written page that cannot be reached scores zero, and an afternoon spent rewriting titles on pages nobody can crawl is an afternoon you will never get back.
Pass two, is the right page answering the right query
The second pass is about assignment, and it is where most of the recoverable performance on an established site is sitting.
Sites that grew by addition end up with several pages competing to answer one question. Google describes the mechanism without drama. “Some websites show the same content under different URLs, which is called duplicate content. Search engines choose a single URL (the canonical URL) to show users, per piece of content.”
Worth quoting the next line too, because audits routinely get this wrong and flag it as a penalty. “Having duplicate content on your site is not a violation of our spam policies, but it can be a bad user experience and search engines might waste crawling resources on URLs that you don’t even care about.” So the cost is waste and confusion rather than punishment, which changes how urgently you should treat it and what you should tell your stakeholders.
The output of this pass is a map, one primary query per page, with the losing duplicates marked for merging or for a canonical. It is the least glamorous artifact an audit produces and usually the most valuable, because it stops you optimizing two pages against each other for the rest of the year.
Pass three, the page itself
Only now does it make sense to read individual pages, because only now do you know which pages are reachable and which one owns each query.
Google’s own starter guidance is the shortest useful checklist here, and it is notably less elaborate than most audit templates. Write descriptive link text, because “this text tells users and Google something about the page you’re linking to”. Use descriptive URLs, since “parts of the URL can be displayed in search results as breadcrumbs”. Think about the words a reader would actually search, which the guidance frames as expecting your readers’ search terms, and notes that experts and beginners reach for different words for the same thing.
What you are checking on each page is whether it answers the query you assigned it in pass two, in the vocabulary the searcher used, within the first screen. Most on-page findings reduce to a page that answers a slightly different question than the one it was built to win, which is a rewrite rather than a tweak.
Resist the urge to score each page out of 100 while you do this. A score compresses a page down to a number and then you lose the finding, which was the specific sentence that needed to change.
Where an AI assistant genuinely helps, and where it cannot
Enough people are now searching for how to run audits with an AI assistant that it deserves a straight answer rather than either half of the usual argument.
What it does well is the middle of the procedure. Hand it a crawl export or a query export and it will classify hundreds of rows into groups faster than you will, spot the pattern that two hundred URLs share, draft the finding text, and turn a spreadsheet into a report a non-specialist can read. That is real time saved on the part of an audit that is genuinely clerical.
What it cannot do is the part people expect. It does not crawl your site, it cannot see your search data, and it has no access to anything you have not handed it. Ask it to audit your site from the name alone and it will produce a competent, generic, entirely unverified document, because every specific in it was inferred rather than observed. The failure is quiet, which is what makes it expensive.
So the rule is the same one you would apply to a junior analyst. It works on data you give it, and every factual claim it returns about your site gets checked against the export it came from before it reaches a report. The useful division is that it should handle volume and you should handle judgment, because the expensive decisions in an audit are about which findings matter on this particular site, and that call needs context no export contains.
Stage three, turning findings into a report somebody acts on
An audit that does not get acted on is a document, and the difference between the two is almost entirely in how the findings are written.
Every row needs four columns and the test is whether a developer could start without asking a question. What is wrong, stated as a condition rather than a category. Where, as a URL or a URL pattern rather than a section name. What it costs, in impressions or clicks at stake rather than in severity words. And what to do, specifically enough to estimate.
The common failure is a report made of categories, where a row says page speed or thin content and the reader has to do the diagnosis again themselves. Traffic lights are the same failure in color. If a row cannot name a URL, it is not a finding yet, it is a topic.
Our guide to what goes in an SEO report covers this from the receiving end, which is a useful way to write one, because the questions a skeptical reader asks are exactly the columns most reports are missing.
How to order the fix list when everything looks urgent
A finished audit is longer than anybody’s capacity to act on it, so the ordering rule does more work than the findings do.
Order by traffic at stake divided by effort, not by severity. Severity is the auditor’s opinion about how wrong something is. Traffic at stake is a number you already exported at baseline stage, which is another reason stage one comes first. A critical issue on a page nobody reaches loses to a moderate issue on the page that earns.
Then apply one filter. Anything that cannot be described as a change to a named URL goes to the bottom, regardless of how important it sounds, because work that cannot be scoped will not be scheduled and will sit on the list making the list look longer.
If the list is being handed to a retainer rather than to your own team, the shape of what you can get through in a month is a separate planning problem, and our breakdown of SEO maintenance packages by tier covers what different levels of commitment actually get through.
Stage four, set the re-measure date before you start fixing
Agree the date in the same meeting where you agree the work, because agreeing it afterward means agreeing it once somebody already knows which date flatters them.
Google is unusually direct about the timing. “Every change you make will take some time to be reflected on Google’s end. Some changes might take effect in a few hours, others could take several months. In general, you likely want to wait a few weeks to assess whether your work had beneficial effects in Google Search results.”
Two things follow. A review held the week after a deploy is too early to conclude anything, and a review held at six months has too many other changes inside it to attribute anything. Pick a date a few weeks out, write it down, and record what else is scheduled to ship in that window so you know in advance what will muddy it.
Whatever you use to pull the comparison, keep it the same instrument as the baseline. Switching tools between the two measurements is the single easiest way to manufacture a result, and our look at what goes in an SEO report dashboard covers what these platforms do and do not hold on to.
Stage five, telling a result apart from noise
The comparison is the easy part. Deciding what it proves is where most of these projects quietly fall apart.
Compare like periods rather than adjacent ones, so the same weekdays and the same season sit on both sides. Compare the pages you changed against the pages you did not, because if both rose by the same amount then something happened to the whole site and it was not your fix. That internal control costs nothing and settles more arguments than any chart.
Then write down what did not move, in the same document, with the same prominence. A performance review listing only improvements is not a review, and the pages that failed to respond are the most useful input into the next audit because they are where your model of the site was wrong.
If you are comparing yourself against other sites rather than against your own history, that is a different measurement with different traps, and our guide to benchmarking SEO side by side handles it.
What this procedure does not cover
Worth naming the edge of this, because the phrase website audit gets used for two genuinely different jobs and buying the wrong one is common.
Everything above audits your site as a search asset. It asks whether pages can be reached, whether the right one answers each query, and whether search behavior changed after you acted. It does not ask whether the site is fast, whether it is usable with a keyboard or a screen reader, or whether it is safe. Those are properties of the site as a delivered product and they cost money in ways that never appear in a query export.
That second job is its own procedure with its own instruments, and our guide to conducting a website audit is where the speed, accessibility and delivery side belongs. The two overlap on exactly one finding, which is that a page nobody can load is also a page nobody can rank, and beyond that they are separate run books that answer to separate people. A useful way to decide which you need is to ask who would receive the report. If the answer is whoever owns the site’s content and its search results, you want this one. If it is whoever owns the build, the hosting and the front end, you want the other.
Scoring a site overall, as a single grade across both, is a third thing again and mostly a reporting convenience. Our look at what an SEO score is made of covers what those composite grades are actually built from.
What we would check first
If you have an afternoon rather than a two weeks, the procedure compresses without breaking, as long as you keep the order.
Export the baseline first, even though nothing is broken yet, because it is the only step that expires. Then run pass one against your twenty highest-earning URLs rather than the whole site. Most sites of any age have at least one page in that twenty that is not reachable the way its owner believes it is, and finding that on day one usually pays for the exercise.
Then write the re-measure date in the calendar before you fix a single thing. It takes ten seconds and it is the difference between a project that can be judged and one that will be argued about.
If the crawl and indexing half is where it falls down, that is the part with the most specialized tooling and the least tolerance for guesswork, and it is what our technical SEO audit covers end to end, with the baseline taken before anything changes for exactly the reasons above.
