A scanned conference figure may be the largest file on a research page—and the clearest evidence for its argument. Deleting it saves bytes but weakens the article. Instead, load a smaller display copy with the text and keep the original available for close inspection. Preserving a source need not mean sending its full-resolution scan to every visitor.
Research-history pages make that distinction especially important. A post about mid-2000s speech recognition may need a spectrogram, error-rate tables, citations, and an account of the test corpus. A computer-vision post might need several frames to show where tracking failed under occlusion. Strip those out and the page becomes faster but less informative. The aim is to cut unnecessary transfer and delay without cutting the evidence.
Measure the experience before changing the page
“Page speed” is not one number. The server may be slow to respond, the browser may be waiting for a large image, or the page may look ready while a script interrupts taps and scrolling. Test a representative article on a phone-sized viewport and a slower connection, not just on the desktop used to edit it. Note which element dominates the first screen, which files account for most of the transfer, and whether the text is usable before the last asset arrives.
Different measurements point to different faults. Largest Contentful Paint marks when the largest visible content element appears. Interaction to Next Paint measures responsiveness after a user interacts with the page. Cumulative Layout Shift records unexpected movement as elements load or change size. None tells you whether the history is sound. A good speed score is little comfort if the crucial scan cannot be read.
Test more than the homepage. A short text post, a long illustrated article, and a document-heavy page may behave quite differently under the same theme. If real-visitor measurements are available, compare them with laboratory tests: a lab run can reproduce a fault, while visitor data captures varied devices and connections. Check first visits separately from return visits, since browser caching can hide the weight of an uncached page.
- Record a baseline: note page size, major requests, prominent delays, and visible layout movement.
- Find the cause: identify the image, font, script, stylesheet, or server delay involved rather than blaming all images.
- Change one category at a time: otherwise, a media improvement may be confused with a script regression.
- Retest the article: make sure captions, citations, diagrams, and controls still work.
Keep evidence, but deliver the right version first
An archival master and a web display image have different jobs. The master holds detail for inspection or future reuse; the display copy needs to suit its place in the article. Sending a multi-megabyte scan to fill a 700-pixel-wide figure adds transfer without adding visible detail at that size. Make a suitably sized copy and provide the larger file when readers need to inspect fine print.
Resolution is only part of the decision. A photograph, a scanned page of text, and a diagram with sharp labels respond differently to compression. Inspect each result at its published size and close up. If the axis labels on a stereo-reconstruction plot become illegible, the smaller file has cost you evidence. Modern image formats may reduce transfer substantially, but check the site's fallback behavior and look closely at fine lines, text, and tonal detail after conversion.
Use responsive image variants so a narrow screen does not fetch a desktop-width file. Set explicit width and height, or otherwise reserve the image's space, so the paragraph below does not jump when it loads. Figures farther down the page can often load as the reader approaches them. Do not automatically defer the main figure at the top: if it is the largest visible element, delaying its request can make the page feel slower.
Captions and surrounding prose carry part of the evidence. A photograph of a laboratory setup needs context so readers can tell original apparatus from a later illustration. For a graph, explain its axes and comparison conditions in the text rather than leaving the only record in tiny embedded labels. That context also helps readers decide whether they need to open the full-size file.
Choose a format suited to the material
Line diagrams with labels may stay sharp as vectors, provided you check font and rendering behavior. Screenshots and scanned documents often call for careful compression rather than aggressive photographic settings. For object tracking, a short sequence of frames may be more useful than an autoplaying clip: the author can point to the exact frame where a target disappears, and the reader need not download video on arrival.
If motion is essential—say, to show a tracker recovering after occlusion—start with a static preview and a clearly labeled play control. Add a concise account of what the clip demonstrates. Readers should be able to follow the argument without an unrequested media transfer, while still having access to the motion for closer study.
Make documents accessible without forcing a download
Historical technical articles often refer to scanned papers, conference programs, or tables preserved as PDFs. Embedding a whole document in every article can impose that cost before anyone chooses to inspect it. Give it a descriptive label and enough bibliographic context to show why it matters, then let readers open or download it deliberately. For a central excerpt, a legible image paired with an accurate transcription may work better than a cramped PDF viewer.
Transcribe carefully. Preserve distinctions that affect interpretation, including dataset names, units, and qualifications in captions. Mark uncertain words instead of guessing silently, and say when a reproduction is incomplete. Searchable text is more than a speed workaround: it lets readers find and assess evidence that may be difficult to read in a scan on a small screen.
Apply the same judgment to long tables. If ten rows establish the comparison discussed in the paragraph, an HTML table containing those rows may be more useful than an image of the full paper page. Keep headings, units, and explanatory notes with the data, and leave the original available for readers who need the wider context.
Remove work that contributes nothing to the article
Images draw attention, but unrelated code may be the bottleneck. A theme can load styles for components the article never uses. A video embed may request outside scripts before anyone presses play. A font family may arrive in several weights when the page uses only two. Each request may have a purpose somewhere on the site; it does not necessarily belong on every post.
Audit scripts by function. Keep navigation, image viewing, and search features that visitors need, but do not make article text wait for optional widgets. Load nonessential scripts without blocking the first render, then check that delayed execution has not broken keyboard controls or caused a sudden layout shift. Verify that a feature is unused before removing it; an unfamiliar script should not be disabled merely because a report flags it.
Fonts need a readability check as well as a transfer check. A system font avoids a download, though a well-configured web font may suit a publication's typography. Limit unused weights and character sets where practical, choose sensible fallbacks, and watch for late swaps that move headings or table columns. Small labels inside research figures should not depend on the article font loading.
For a research archive, this belongs alongside routine upkeep. The post on website maintenance for a historical research archive covers the broader work; a speed audit asks specifically what each page transfers and when.
Improve delivery before cutting explanations
If the initial response is slow, a smaller figure may do nothing for the delay before content appears. Investigate server response time, caching, database-heavy page generation, and unnecessary redirects. Caching can spare the server from rebuilding a frequently requested article for every visitor, but it needs a reliable update process: corrected captions and revised citations must reach readers.
Compression reduces the transfer cost of text-based HTML, CSS, and JavaScript. Browser caching saves repeat visitors from fetching unchanged assets again. Neither replaces editorial judgment. After changing cache settings, test a published correction in a fresh browsing session as well as a logged-in editor session; editors may see an updated page while readers still see a cached one.
Keep the rendered page straightforward. Semantic headings, paragraphs, lists, and tables can carry explanations without extra interactive machinery. If a post compares several mid-2000s language-identification systems, a clear table and a short note on differing test conditions may say more than an animated chart. Interactivity can earn its place, but it should add something to the evidence.
Set priorities by reader value and loading cost
The choice is rarely just “keep or delete.” For each heavy element, ask what question it answers, whether the page answers that question elsewhere, and when the reader needs it. Those answers suggest ways to reduce cost without thinning the account:
| Page element | Why it matters | Lower-cost delivery |
|---|---|---|
| Figure central to a claim | Allows readers to inspect the reported result | Legible display copy with access to a larger version |
| Several supplementary scans | Document details beyond the main argument | Smaller previews loaded later, with descriptive labels |
| Short demonstration video | Shows motion a still cannot convey | Static preview; load the player when requested |
| Long bibliography | Enables source checking | Keep as text rather than an image or embedded document |
| Decorative background image | Adds atmosphere but no evidence | Compress heavily or remove |
Historical sources make these trade-offs especially visible. A reproduced conference poster may contain a valuable architecture diagram alongside columns of text that become illegible at phone width. A tiny poster preview is not enough. Show the diagram at a readable size, explain how it relates to the paper, and offer the complete poster separately if the site has permission to reproduce it. Readers can get the argument promptly and still inspect its source.
Check what optimization changed for a real reader
After each change, compare the same pages under similar test conditions—and inspect them yourself. A speed gain matters only if the article remains stable and understandable. On a small screen, check for clipped equations, unreadable figure labels, missing table columns, and captions stranded away from their images. Open the full-size scan and return to the article. If you added a preview or expandable section, try it with a keyboard.
Read a paragraph without its media, too. If it says “as the figure shows” but never states the observation, the prose is asking the file to do too much. On the other hand, when a figure documents a difference that words cannot faithfully convey, keep a legible version rather than substituting a vague description. A faster asset will not fix an unclear explanation.
For a first pass, choose one image-heavy article and list its three largest transferred files. Suppose the biggest is a scanned chart displayed at 600 pixels wide. Make a readable 600-pixel display copy, retain the original for inspection, and check both the transfer size and the smallest chart label on a phone before publishing.