Blog

Short answer: the standard wordpress database cleanup check for autoloaded options, WHERE autoload = 'yes', returns nothing on current WordPress. The column now holds on, off, auto and the legacy no. A query matching zero rows looks identical to a database with nothing to clean, which is how sites end up carrying megabytes nobody has measured.

I found this by running the check on a stock install and getting a flat NULL back.

The values actually in the column

older versions carry no as well, so a correct query has to handle both eras.*

Value	Rows	Meaning
on	213 (66.0KB)	explicitly autoloaded
auto	84	left to WordPress, which does autoload these
off	136	explicitly not autoloaded
no	present	legacy value, still found on upgraded sites
yes	0	every tutorial’s query target, matches nothing on a fresh install

Caveat: one install, WordPress 7.1. The column values themselves are a core schema change, documented and verifiable on any current install, not a one-off finding. The row counts are specific to this site and will differ on yours; the query correction is not.

Two of those values mean autoloaded. on is explicit, and auto leaves the decision to WordPress, which does load them. So a real measurement matches on, auto and legacy yes together.

run it again with the corrected clause. Matching nothing looks exactly like nothing to fix.*

Re-running both queries on a real, plugin-heavy install

The original finding came from a near-empty stock install. I re-ran both the broken query and the corrected one against a different, considerably busier install, forty plugins present, WooCommerce, Elementor and Yoast SEO active among them, to see whether a real site’s actual plugin load changes the picture.

SELECT COUNT(*), ROUND(SUM(LENGTH(option_value))/1024,1) FROM wp_options WHERE autoload = 'yes';
SELECT COUNT(*), ROUND(SUM(LENGTH(option_value))/1024,1) FROM wp_options WHERE autoload IN ('on','auto','yes');

autoload = 'yes'                  ->  0 rows,   NULL
autoload IN ('on','auto','yes')   ->  332 rows,  71.9 KB

The broken query still returns exactly nothing on this second, busier install, confirming the schema change is not a one-off oddity of a fresh install specifically. The corrected query finds 332 autoloaded rows totalling 71.9 KB, close to the original reading’s 297 rows and 66 KB despite this install carrying a genuinely larger and more active plugin set. That similarity is itself worth a note: autoload weight does not scale simply with plugin count, because most of what a busy plugin adds to wp_options lands in the off category correctly, 164 rows and 1,080.6 KB of it here, and never touches every request at all.

Reading the top of the corrected list on this install shows exactly the kind of drift a real site accumulates that a bare install cannot demonstrate:

_transient_wp_core_block_css_files   23.22 KB
_transient_wp_styles_for_blocks      11.08 KB
wp_user_roles                        10.85 KB
wpseo_titles                          8.42 KB
cron                                  4.91 KB
wpseo                                 4.75 KB

wpseo_titles and wpseo, together over 13 KB, are Yoast SEO’s own configuration, present here specifically because this install has an SEO plugin active, absent from the original post’s near-empty baseline entirely. That is the practical shape of autoload growth on a real client site: not one runaway option, but each plugin adding its own few kilobytes, none individually alarming, all of it read on every single request regardless of whether the page being served needs any of it.

Why autoload is the number that matters

Every autoloaded option is read from the database on every single request, including requests that need none of it. It is loaded before WordPress knows whether the visitor is asking for the homepage, a REST endpoint, or an image.

reason to know the number is that it only ever grows, and nothing tells you.*

Note the top of that list: two of the three largest rows are transients that are autoloaded. A transient is a cache with an expiry, and these are being read on every request whether or not anything needs them. That is normal core behaviour rather than a bug, and it is why a site with years of plugin history routinely reaches several megabytes.

The cron row is worth a glance too. That single option holds the entire scheduled-task array, which means a site whose scheduler has quietly stopped is still paying to load its backlog on every page view.

On this install, options beat postmeta, and that is worth explaining rather than hiding

Running the full table-size query this post recommends against the same lab install produced a result that contradicts this post’s own claim about which table usually dominates, and it is worth showing rather than quietly adjusting the earlier text to match.

SELECT table_name, ROUND((data_length + index_length)/1024/1024, 2) AS mb, table_rows
FROM information_schema.tables WHERE table_schema = DATABASE()
ORDER BY (data_length + index_length) DESC LIMIT 8;

wp_options                    2.14 MB   491 rows
wp_actionscheduler_actions    0.16 MB    17 rows
wp_wc_orders                  0.16 MB     0 rows
wp_comments                   0.14 MB     3 rows
wp_posts                      0.14 MB    32 rows

wp_options at 2.14 MB, wp_postmeta nowhere in the top five at all, a mere 0.05 MB across 107 rows. This post’s earlier section calls wp_postmeta “almost always the biggest,” and on this specific install that is simply not true.

The reason is install age, not a flaw in the general claim. wp_postmeta grows with content volume and history, every post, every revision, every piece of metadata a plugin ever attached, accumulated over years. This lab install has 32 posts and no real editorial history behind it. wp_options, by contrast, front-loads a large amount of its weight the moment plugins are activated, page-builder configuration, SEO plugin settings, WooCommerce’s own setup data, regardless of how much content exists yet. A young, plugin-heavy site can easily carry more weight in wp_options than in wp_postmeta for exactly that reason, and the general claim this post makes describes the steady state a site reaches after years of content, not the shape of a site in its first months. Running the actual query, rather than trusting either claim by default, is the only way to know which shape a specific site is actually in, which is the entire argument this post has been making from its first paragraph, now proven against itself.

The tables that actually grow

Autoload is the number that affects every request, but it is rarely the largest thing in the database. Four tables do most of the growing on a long-lived site.

wp_postmeta is almost always the biggest. Every plugin that attaches data to posts writes here, and a plugin removed two years ago usually left its rows behind. Orphaned rows, where the parent post no longer exists, are safe to delete once you have verified the join.

wp_options grows in two directions: autoloaded rows that cost you on every request, and expired transients that cost you storage. Both are worth clearing, for different reasons.

wp_comments and wp_commentmeta on any site that has ever had comments open. Spam that was caught and never emptied sits there indefinitely.

Plugin tables. Form submissions, order line items, activity logs, licence checks. These are the ones that reach gigabytes, and they are the ones a generic cleanup tool must never touch.

Measure before assuming:

SELECT table_name,
       ROUND((data_length + index_length)/1024/1024, 1) AS mb,
       table_rows
FROM information_schema.tables
WHERE table_schema = DATABASE()
ORDER BY (data_length + index_length) DESC
LIMIT 12;

That single query tells you exactly where the size actually is, which is frequently not where you expected it to be.

What is safe to remove

tables it does not recognise is offering to delete a client’s orders.*

Target	Safe?	Notes
expired transients	yes	a cache by definition, WordPress regenerates them
auto-drafts and trashed posts	yes	respect the trash retention period first
spam and trashed comments	yes
orphaned postmeta and termmeta	yes	verify the join before deleting
post revisions	ask first	some clients genuinely use them
autoloaded options from removed plugins	check first	set autoload to off rather than delete
anything in a plugin’s own tables	no	order data, form entries, licence records

Two judgement calls worth making deliberately.

Revisions. Purging them is the default advice and it is wrong for some clients. Editors who actually use revisions will notice. Cap them going forward with WP_POST_REVISIONS instead of deleting history somebody may rely on.

Options from removed plugins. Set autoload to off rather than deleting the row. The plugin may come back, and an option that is not autoloaded costs nothing on requests that do not read it.

Running a WordPress database cleanup without breaking anything

The order matters more than the technique.

Take a backup and prove it restores. Not take one. Prove it, because a cleanup is exactly the operation you might need to undo.

Measure before. Record the autoload total and the ten largest rows. Without a before number you cannot tell a client what changed.

Change one category at a time. Transients, then trash, then orphaned meta. If something breaks you want to know which step did it.

Measure after, and load the site. Homepage, a form, the checkout if there is one.

On most sites this recovers less than people expect, and that is worth saying plainly. Database cleanup is not usually the reason a site is slow. It is worth doing because unbounded growth eventually becomes a real problem, not because it will transform a page load next Tuesday.

If the site genuinely is slow, start with where the time is actually going rather than with the database, or you will spend an afternoon optimising something that was never the bottleneck.

Protuno’s Performance agent, Dash, carries a Database Optimisation playbook. It can change a site, so it waits for the owner’s approval in Protuno and takes a backup first. Dash is live now.

One habit worth adopting, because it turns this from a one-off tidy into something useful. Record the autoload total and the top three tables in the client’s notes each quarter, with the date. Four numbers, thirty seconds. The value is not any single reading; it is that a year later you can see the shape of the growth and say something specific, like this table has tripled since March and here is what changed then. Nobody can act on “the database is quite big”. Everybody can act on a trend with a date attached.

Run the corrected autoload query on your largest client site today. If it comes back over a megabyte, you have found something worth an afternoon.

Run the table-size query on the same site immediately afterwards, rather than stop at autoload alone. The two queries answer different questions, one measures the cost paid on every request, the other measures where the disk space actually sits, and a site can score badly on either one independently of the other. Checking both, on the actual site in front of you rather than assuming which one matters more from a general rule, is the only way to know which afternoon is actually worth spending.

Seeing the autoloaded-options problem on a real site

A practical walkthrough of the exact Site Health warning this post’s autoload query is built to catch, including why the warning does not update immediately after a fix ships.

Short answer: wordpress image optimization starts with measuring what a page actually downloads, not with installing a plugin. Fetch every image the page references, sort by bytes, and look at the top three. On one of our own pages that was 2,590 KB across forty images, all of them PNG.

What the audit found

the dates: images uploaded in 2021 are still being served on a 2026 page.*

That last point is the part people miss. A media library is append-only in practice. Nobody revisits a decorative image from four years ago, and nothing flags it, so it keeps being served long after the page it belongs to was redesigned around it.

Running a WordPress image optimization audit yourself

You do not need a tool. The method is: read the page, collect every img src, source srcset and CSS url(), request each one, and sort by response size.

curl -s https://example.com/ 
  | grep -oE 'src="[^"]+.(jpe?g|png|webp|avif)"' 
  | sed 's/src="//;s/"$//' | sort -u 
  | while read -r u; do
      printf '%8s  %sn' 
        "$(curl -s -o /dev/null -w '%{size_download}' "$u")" "$u"
    done | sort -rn | head

Two refinements that matter in practice.

Request as a browser would. Send an Accept header including image/webp and image/avif. Some servers negotiate format, and requesting with curl’s default headers will show you the fallback rather than what visitors actually receive.

Follow srcset, not just src. Responsive images mean the file a phone downloads is often not the one in src. Audit the largest candidate, because that is what a desktop visitor gets.

Checked again, and the page got heavier, not lighter

I re-ran this post’s own audit against the same sitea.example homepage before expanding it, expecting either no change or a modest improvement. Neither happened.

curl -s https://sitea.example/ | grep -oE 'src="[^"]+.(jpe?g|png|webp|avif)"' 
  | sed 's/src="//;s/"$//' | sort -u | while read -r u; do
    curl -s -o /dev/null -w '%{size_download} %{url_effective}n' "$u"
  done | sort -rn

	Original reading	This reading
Images on the page	40	64
Total image weight	2,590 KB	6,475 KB
Largest single image	794 KB (00-feature-image.png)	838 KB (a new file, ...-Elmentor-Email.png)

Sixty-four images now, up from forty, and total weight up to 6,475 KB, nearly two and a half times the original figure. The original 813,352-byte file is still there, unchanged, now sitting at the second-heaviest position rather than the first, because a new 857,815-byte PNG, uploaded in June, has since overtaken it. Nothing about the original finding was fixed. The page simply grew a new, larger problem on top of the old one that was never addressed.

I checked the new heaviest file the same way this post’s format-negotiation table checks the old one:

curl -s -H 'Accept: image/avif,image/webp,image/png,image/*' -D - -o /dev/null 
  "https://sitea.example/wp-content/uploads/2026/06/site-notification-email.png" 
  | grep -i content-type

content-type: image/png

Same pattern as before: the Accept header is ignored regardless of what it asks for, and .webp returns 404 when requested directly. Whatever caused the original file to bypass format negotiation is still causing every new upload to do the same thing, which means this is not a one-off oversized file. It is a standing gap in how images reach this page at all, quietly accepting new casualties every time someone adds a graphic.

Format is only half of it

more image weight, including a single 1.4 MB WebP.*

	sitea.example	woocommerce.com
Images on the page	40	40
Total image weight	2,590KB	6,418KB
Format served	PNG throughout	WebP throughout
Largest single image	794KB (png)	1,412KB (webp)

Caveat: one page from each site, checked once. This is not a claim that WebP sites are heavier in general, it is the opposite point: format alone does not bound page weight, and this pair happens to prove it cleanly. Do not skip the dimension check because a site already converted formats.

Converting to a modern format shrinks a file. It does not stop somebody uploading a photograph at twice the dimensions the layout will ever display, and it does not remove an image the design no longer uses.

is no WebP sibling to fall back to. woocommerce.com serves WebP from URLs that still end in .jpg, because something in front converts per request.*

Request	sitea.example
Accept: image/png	813,352 bytes, image/png
Accept: image/webp,image/*	813,352 bytes, image/png (ignored)
Accept: image/avif,image/webp,...	813,352 bytes, image/png (ignored)
Does a .webp sibling exist?	404, no

woocommerce.com, by contrast, serves image/webp from URLs that still end in .jpg and .png, because something in front of it converts per request regardless of the extension.

So the audit has three questions, in this order:

Is this image needed at all? The cheapest image is the one you removed.

Are the dimensions sane? A 3000px-wide file displayed in a 600px column carries twenty-five times the pixels the layout can use, and no format conversion recovers that.

Is the format modern? This is the last question, not the first, and it is covered separately in the piece on WebP and AVIF, publishing later this week.

The comparison this post opened with has nearly closed

The original table in this post used woocommerce.com as the counterexample: a WebP site still shipping more image weight than a PNG one, proof that format alone does not bound page weight. I re-checked woocommerce.com the same way, and the comparison this post built itself around looks different now.

curl -s https://woocommerce.com/ | grep -oE 'src="[^"]+.(jpe?g|png|webp|avif)"' 
  | sed 's/src="//;s/"$//' | sort -u | while read -r u; do
    curl -s -o /dev/null -w '%{size_download}n' "$u"
  done | awk '{s+=$1} END {print s}'

	sitea.example (PNG)	woocommerce.com (WebP)
Images	64	57
Total weight	6,475 KB	6,637 KB

162 KB apart, on 6.5 megabytes of images. What was originally a clean illustration, PNG site lighter in file count but present anyway, WebP site heavier despite doing format conversion correctly, has become two sites of nearly identical total weight, arrived at by different routes: one added twenty-four new images without converting any of them to a modern format, the other apparently kept pace on its own. The lesson this post opened with, format is not the first question and page weight can grow regardless of it, holds more strongly now than it did originally, not less, because the specific numbers that used to make one site look like the counterexample have converged into two sites making the same mistake for different reasons.

Where the weight actually concentrates

Worth naming precisely rather than leaving as “many images add up”: on the current sitea.example homepage, the three heaviest files alone, 858 KB, 813 KB and 542 KB, total 2,213 KB, just over a third of the entire page’s 6,475 KB of images, out of sixty-four files. Three files, out of sixty-four, carrying more weight than the other sixty-one combined would if evenly distributed.

That concentration is exactly why this post recommends sorting by size and starting at the top rather than working through a full image inventory in upload order. Fixing the three heaviest files on this specific page would remove more weight than optimising the next twenty smaller ones, and it is a single afternoon’s work rather than a full site audit. The audit loop this post already gives sorts for exactly this reason: the answer to “where do I start” is almost always sitting in the first three lines of output, not buried somewhere in the middle of a sorted list nobody scrolls to.

Where this fits

Image weight is the most common single cause of a slow-feeling page, and it is also the one that degrades silently. Nobody ships a 794 KB image deliberately. Somebody exported a mockup at full resolution once, and every visitor since has paid for it.

It is worth checking alongside whether compression is even reaching your assets, because an uncompressed JavaScript bundle and an oversized hero image are usually found on the same site, for the same reason: nobody has measured.

And measure the page, not the server. A fast server response time with three megabytes of images behind it is a fast server attached to a slow page.

Protuno’s free audit flags oversized images from the domain alone as part of the performance checks. Dash is live now.

One organisational fix is worth more than any plugin here: put a ceiling on what can be uploaded. Most oversized images arrive because somebody exported at full resolution and WordPress accepted it without comment. A maximum upload dimension, agreed with whoever writes the content, stops the problem at the source rather than compressing it afterwards. Optimising a 4000px image that should never have been 4000px is solving the wrong problem well.

Run the loop above on your client’s homepage. The top three lines are your afternoon’s work, and they are usually worth more than everything else on a performance checklist combined.

Why the same page kept growing without anyone deciding it should

Nobody sat down and decided sitea.example‘s homepage should carry six and a half megabytes of images. Twenty-four images arrived one at a time, each added for a real reason, a new section, a new screenshot, a new testimonial logo, by different people over different months, none of whom were looking at the running total the way this audit does. That is worth naming as the actual mechanism behind this post’s second reading, not a one-off oversight but the default outcome of a page that anyone with publishing access can add an image to, with nothing in the workflow surfacing what the page now weighs as a whole.

That is also why re-running this exact check periodically matters more than running it once thoroughly. A single audit, however careful, describes the page on the day it ran. The next image added the following week is invisible to it. A page that never gets re-measured will keep growing in exactly the direction this one did, because every individual addition looks small and reasonable in isolation, and the only view that shows the accumulated cost is the one this post keeps insisting on: fetch everything, sum it, sort it, and look at what the page actually weighs today rather than what it weighed the last time anyone checked.

What would have caught this before it doubled

The upload-ceiling recommendation later in this post, a maximum dimension enforced at the point of upload, would not have caught this specific growth, because the problem here is not that individual images are oversized for their display dimensions. It is that the count of images grew without anyone tracking the running total. A complementary check worth adding alongside the dimension ceiling: alert when a page’s total measured image weight crosses a threshold, checked the same way this post’s loop already checks it, rather than only checking each new image in isolation as it is uploaded. A single 400 KB image added to a page that already carries six megabytes is a smaller problem than the same 400 KB image added to a page that started at zero, and a per-image check alone cannot see that difference. Only a running total, measured against the whole page, catches death by a thousand reasonable-looking additions.

The other side of the same problem: images nobody is even using

Oversized images that are still in use are this post’s subject. Unused images sitting in the media library, never referenced on a live page at all, are the adjacent failure this walkthrough covers, and they cost real storage and backup time even though no visitor ever loads them.